Special Issue on Artificial Intelligence in Economics, Finance and Business
- 35 -
* Corresponding author.
E-mail address: [email protected]
Keywords
Class Imbalance, Down-
Sampling, Ensemble
Approaches, Machine
Learning, N-grams,
Over-Sampling
Techniques, TFIDF.
Abstract
Online portals provide an enormous amount of news articles every day. Over the years, numerous studies have
concluded that news events have a signiicant impact on forecasting and interpreting the movement of stock prices.
The creation of a framework for storing news-articles and collecting information for speciic domains is an important
and untested problem for the Indian stock market. When online news portals produce inancial news articles about
many subjects simultaneously, inding news articles that are important to the speciic domain is nontrivial. A critical
component of the aforementioned system should, therefore, include one module for extracting and storing news
articles, and another module for classifying these text documents into a speciic domain(s). In the current study,
we have performed extensive experiments to classify the inancial news articles into the predeined four classes
Banking, Non-Banking, Governmental, and Global. The idea of multi-class classiication was to extract the Banking
news and its most correlated news articles from the pool of inancial news articles scraped from various web news
portals. The news articles divided into the mentioned classes were imbalanced. Imbalance data is a big dificulty
with most classiier learning algorithms. However, as recent works suggest, class imbalances are not in themselves
a problem, and degradation in performance is often correlated with certain variables relevant to data distribution,
such as the existence in noisy and ambiguous instances in the adjacent class boundaries. A variety of solutions to
addressing data imbalances have been proposed recently, over-sampling, down-sampling, and ensemble approach.
We have presented the various challenges that occur with data imbalances in multiclass classiication and solutions
in dealing with these challenges. The paper has also shown a comparison of the performances of various machine
learning models with imbalanced data and data balances using sampling and ensemble techniques. From the result,
it’s clear that the performance of Random Forest classiier with data balances using the over-sampling technique
SMOTE is best in terms of precision, recall, F-1, and accuracy. From the ensemble classiiers, the Balanced Bagging
classiier has shown similar results as of the Random Forest classiier with SMOTE. Random forest classiier's
accuracy, however, was 100% and it was 99% with the Balanced Bagging classiier.
DOI: 10.9781/ijimai.2022.02.002
A Comparative Analysis of Machine Learning
Models for Banking News Extraction by Multiclass
Classication With Imbalanced Datasets of Financial
News: Challenges and Solutions
Varun Dogra
1
, Sahil Verma
2
, Kavita
2
, NZ Jhanjhi
3
, Uttam Ghosh
4
, Dac-Nhuong Le
5,6
*
1
School of Computer Science and Engineering, Lovely Professional University (India)
2
Department of Computer Science and Engineering, Chandigarh University, Mohali (India)
3
School of Computer Science and Engineering, Taylor’s University (Malaysia)
4
Department of Computer Science and Data Science, Meharry School of Applied Computational
Sciences, Nashville, TN (USA)
5
School of Computer Science, Duy Tan University, Danang, 550000 (Vietnam)
6
Institute of Research and Development, Duy Tan University, Danang, 550000 (Vietnam)
Received 13 November 2020 | Accepted 19 January 2022 | Published 8 February 2022
I. I
I
N the equity market, stocks or funds belong to the different business
sectors. And sector-based news has become an inseparable part of
the management of inancial assets, with news-driven stock and bond
markets explosively growing. Fund managers take advantage of this
reality and make use of sector-oriented news to select individual stocks
to diversify their investment portfolios to optimize returns. There is
no such structured framework available for classifying the news on
speciic sectors of someone's interest. This problem is increasing by
the day, necessitating a system for news classiication methodology
for speciic sectors.
Machine learning (ML) techniques have demonstrated impressive
performance in the resolution of real-life classiication problems in
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 36 -
many different areas such as inancial markets [1], medical diagnosis
[2], vehicle trafic examination [3], fraud detection [4]. There are
plenty of document classiication systems in the commercial world. For
instance, usually, the news stories are grouped by topics [5], medical
images are tagged by disease categories [6], and many products are
branded according to categories [7]. Different methods of statistical
and machine learning are implemented in text labeling, where one of
the predeined labels is automatically assigned to a given item of the
unlabeled pool of textual articles.
However, the vast majority of articles on the internet about text
classiication are binary text classiication [8] such as email iltering
[9], political preferences [10], sentiment analysis [11], etc. Our real-
world problem is in most cases much more complex than the binary
classiication. More formally, if some d is a document in the whole set
of documents D and C is the set of all categories i.e. C
= {c
1
, c
2
,c
3
,…, c
n
,}
the classiication of text assigns one category c
i
to the document d.
Such a classiication function with more than two classes is known as
multiclass classiication; for example, identify a set of news categories
as business, political, economic or entertainment.
In our paper, we’re interested in isolating news on the banking
sector and its most associated domains from the pool of inancial
news articles. We feel that ‘banking news’ of any nation is most
correlated with their ‘governmental news-events ’ that covers news
on government initiatives for good governance, state or national
elections, change or new development of governmental policies, and
‘global’ inancial news that covers global trade, changes in currency-
commodities prices, and global sentiments. So, we have a 4-class
classiication problem of a set of news articles to extract banking,
and its most correlated news i.e. Government, and Global from entire
inancial news articles. We decide to label the news articles into
banking, governmental, global, and non-banking classes with a total of
10000 instances. The non-banking news covers all the inancial news
scrapped from various new portals divergent from these 3 categories
(banking, governmental and global). The news reports on different
categories are usually imbalanced. The distribution of the news articles
in our dataset is shown in Fig. 1. The news articles are manually labeled
into these four classes. [12] mentions that labeling is normally done
manually by human experts (or users), which is a time-consuming and
labor-intensive process but it results in higher accuracy due to expert
knowledge being involved in labeling text articles with appropriate.
In the process, we label a set of representative news articles for each
class. The labelers are experts in the inancial domain and inancial
markets. A team of three experts is used to perform feature selection
to identify important or representative words for each class used in
a 4-class classiication, followed by inspecting each text document
and label it to the respective class based on representative words for
each class. An agreement is made with the experts to label the given
instances of the news articles. The process is used to derive a set of
documents from entire unlabeled documents for each class to form the
initial training package. The different machine learning techniques are
then applied to build and compare the classiiers. The whole process is
explained in the later part of the paper in sections 3-4.
A. Multiclass Classiication
For machine learning, the problem of classifying instances into
three or more classes in multiclass classiication. Although some
classiication algorithms of course allow the use of more than two
classes, some are by deinition binary algorithms; however, a variety
of strategies may transform these into multi-classiication. In a
multiclass classiication problem, some classes may be represented
with only a few samples (called the minority class), and the rest falls
into the other class (called the majority class). The data disparity in
machine learning creates dificulties in conducting data analytics in
virtually all ields of real-world problems. The problem of classifying
textual news articles is a two-step process. In our experiment, in the
irst step, the documents are collected from various websites like
Bloomberg, Financial Express, and Moneycontrol using web scrapping
code written in Python. It is followed by partitioned news articles
into their respective category of banking, non-banking, global, and
governmental using manual labeling. In the next step, the news articles
are trained and tested using machine learning approaches to achieve
the classiication goal for a new sample of news articles. A comparative
analysis is performed based on the results of the experiment to rate
the tested machine learning algorithms in descending order so they
can be used to evaluate news classiication tasks with imbalanced
datasets. We are not detailing the process of downloading news from
the various sources in the paper.
In turn, multiclass classiication can be divided into three groups:
• Native classiiers: These include most common classiiers such as
Support Vector Machines (SVM), Classiication and Regression
Trees (CART), KNN, Naïve Bayes (NB), and multi-layer output
nodes i.e. Neural Nets.
• Multi-class wrappers: These hybrid classiiers reduce the problem
to smaller chunks that can then be solved with different binary
classiiers.
• Hierarchical Classiiers: Using a tree-based architecture this group
uses hierarchical methods to partition output space into target
class nodes.
B. Learning From Imbalanced Dataset
A dataset is considered class-imbalanced if the number of examples
that represent each class is not equal. Dealing with an imbalanced
dataset has been a popular subject in the research study of classifying
news articles. The conventional machine learning algorithms may
introduce biases while dealing with imbalanced datasets [1]. The
accuracy of many classiication algorithms is considered to suffer
from imbalances in the data (i.e. when the distribution of the examples
is signiicantly distorted across classes) [13]. Most binary text
classiication applications are of this kind, with the negative examples
far outnumbered positive examples of the class of interest [2]. Many
classiiers assume that examples are evenly distributed among classes
Category
% of Total Number of Records
% of Total Number of Records for each Category. Color shows details about Category.
Category
Banking
Banking
0%
10%
9.16%
16.09%
70.54%
4.21%
20%
30%
40%
50%
60%
70%
Global
Global
Governmental
Governmental
Non-Banking
Non-Banking
Fig. 1. The distribution of news article instances amongst 4-classes.
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 37 -
and assume an equal cost of misclassiication. For example, someone
works in an organization and is asked to create a model that predicts
whether news belongs to class A, based on the distribution of news in
classes A and B at your side. He chooses to use his favorite classiier,
train it on data, and before he knows it, he gets an accuracy of 95%.
Without further testing, he wants to use the model. A couple of days
later he underlines the model’s uselessness. Indeed, from the time it
was used to gather news, the model he created did not ind any news
belonging to class A. He igures out after some investigations that
there is only about 5 percent of the news produced in the pool that
belongs to Class A and that the model always responds to Class “B,”
resulting in 95 percent accuracy. The kind of “guileless” indings that
he obtained were due to the imbalanced dataset with which he works.
The goal of this paper is to examine the various methods that can be
used with imbalanced groups to tackle classiication problems.
In the imbalanced data set, basically with this problem, a classiier’s
output leans to be partial towards certain classes (majority class)
[14]. In Natural Language Processing ( NLP) and Machine Learning
in general, the problems of imbalanced classiication, under which
the number of items in each class for a classiication process differs
extensively and the capacity to generalize on dissimilar data remained
critical issues [15]. Most classiication data set do not have precisely
the same number of instances in each class but a slight variation is
often insigniicant. There are problems where class inequality is
believed to not just normal.
Also, classiiers are typically built to optimize precision, which in
the situation of imbalanced training data is not a reasonable metric
for determining effectiveness. Therefore, we are presenting the
comparison of various machine learning classiication techniques
which might result in high accuracy even with imbalanced datasets,
however, it is worth mentioning certain challenges we ind to deal
with imbalanced data and evaluating certain measures along with
accuracy to evaluate performance. Also, we conduct machine learning
on documents to perform multi-classiication, where the data sample
belongs to one of the multiple categories exactly.
The readers will also come to know the following key points after
they have studied this paper:
• Imbalanced classiication is the classiication issue when the
training dataset has an uneven distribution of the classes. As a
result, appropriate sampling techniques must be implemented to
balance the distribution by taking into consideration the various
characteristics and the balanced performance of all of them.
• The class distribution imbalance may vary, but a serious imbalance
is more dificult for modeling and may require advanced techniques.
It is possible to introduce an eficient hybrid ensemble classiier
architecture that incorporates density-based under-sampling or
over-sampling and cost-effective methods by examining state-of-
the-art solutions using a multi-objective optimization algorithm.
• Most real-world classiication problems, such as scam detection,
news headlines categorization, and churn prediction, have an
imbalanced class distribution. Certain issues should be addressed
when constructing multi-class classiiers in the case of class
imbalances.
The paper’s structure is as follows. In Section 2 we present a review
of several current literature methods that handle the classiication of
imbalanced datasets for text classiication. In section 3 we present
our framework of classifying news articles along with challenges
and possible solutions for the classiication of imbalanced datasets.
Sections 4 presents the comparative study of different techniques
along with the experimental outcomes. Section 5 summarizes the
paper and presents the future direction in the area of classiication of
imbalanced datasets.
II. L R
We will present the necessary review in text classiication and
imbalanced learning in the subsequent subsections. We also assess the
state-of-art research involving both the learning of imbalances and
multiclass text classiication.
A. Machine Learning for Text Classiication
Here, we present the relevant literature work in the area of text
classiication using approaches to machine learning. Most of the
preceding research had effective results using supervised methods of
learning [7], [9], [16]. The following sub-sections present the literature
work on feature extraction, selection, representation, and classiication
using learning models.
1. Document Representation
The eficiency of machine learning approaches largely depends
on the option of representation of the data on which they would
be implemented. For this purpose, most of the practical work in
implementing machine learning algorithms runs further into the
creation of pre-processing pathways and data conversion that
leads to the representation of data that can help eficient machine
learning. These representations or attribute development is essential,
yet labor-intensive, and illustrates the vulnerability of current
learning algorithms: their weakness to isolate and arrange the
data discriminatively. However, the goal is clear when it comes to
classiication; we want to reduce the number of misclassiications
upon testing data and overcoming the mentioned challenges in our
framework.
Several machine learning implementations within the text
ield use bag-of-words representation where terms are deined as
dimensions with word frequencies corresponding values. Normalized
representation of the word frequencies is used by many applications
as the dimensional values. One of the signiicant techniques of
describing a document is Bag of Word (BoW). Use the frequency count
of every term throughout the text, the BoW is used to form a vector
describing document. This method of representation of documents is
called a Vector Space Model [17]. However, the relative frequencies
of terms often vary widely, which contributes to the differential
meaning of the different words in classiication applications [18].
With the varying lengths of various text documents, one needs to
normalize when measuring distances between them. To solve these
issues, term weighting methods are used to assign correct weights
to the word for improving text classiication eficiency [19]. Term
weighting has long been developed in machine learning in the form
of term frequency times inverse document frequency i.e. tidf [20].
[21] suggests techniques to improve the TF-IDF scores to improve the
representation of the term spreading between classes. Such practices
may be used in various services where bag-of-word-based TF-IDF
features are used. Equation (1) is given as:
(1)
Here, N represents the overall number of documents and N(t
i
)
denotes the number of documents in which the term ti
occurs in the
collection of documents. tf(t
i
, d
j
), it represents the number of times
term ti occurs in document dj. The newer version is mentioned in (2):
(2)
|T| represents the unique terms available in the collection of
documents,
(3)
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 38 -
The outline in (2) is concerned with the words that belong to
document d
j
.
The importance of the standard term weighting outlines in (1), (2)
is that three basic principles of word frequency distribution have been
integrated into a pool of documents.
1. No less important are uncommon terms than a regular terms-idf
hypothesis.
2. Numerous presences of a word in a text are no less relevant
compared to the presumption of a single appearance-tf.
3. Long documents are no less necessary for the equivalent amount
of term matching than short documents – the assumption of
normalization.
The big drawback of this model is that it results in a large sparse
matrix, which poses a high-dimensionality problem. The design
of such high-dimensional feature spaces is usually inadequate in
the number of items to represent adequately. The reduction of
dimensionality is therefore a signiicant problem for a variety of
applications. The literature has suggested several methods for the
reduction of dimensionality [3], [22], [23]. For such representations,
for instance, linear support-vector machines are comparatively
effective [24]; whereas other techniques like Decision trees have to
be built and modiied with attention to allow their proper usage [25].
When a decision tree induction method computes a decision tree that
depends very much on arbitrary features of the training examples, and
works well only on trained data, but badly on unknown data, the data
becomes overit. There is a way to reduce the chance of overitting by
choosing the perfect subspace for the function at each node [26]. Cross-
validation is an important prevention method to tackle overitting. We
segment the data into sub-sets k, called folds, for regular k-fold cross-
validation. We then train the algorithm iteratively on folds of k-1, thus
using the remainder of the fold as the test set [27].
In several studies, word n-grams were used effectively [21]. N-gram
feature sets include the usage of feature selection approaches to
obtain correct attributes subsets. Word n-grams contain bag-of-words
(BOWs) and word n-grams in higher-order (e.g. bigrams, trigrams).
[28] uses modiied n-grams by integrating syntactic information on
n-gram relationships. In most document classiication activities, this
n-gram model is implemented, and almost always boosts precision.
This is because the n-gram model allows us to take the sequences of
terms into account, as opposed to what will requireto do just by using
single words (unigrams). Looking into the beneits of the n-grams
feature selection, in this paper, a rich collection of n-gram features that
encompassed several ixed and variable n-gram categories is studied
for classifying textual news articles.
2. Feature Selection
The selection of features serves as a crucial technique for reducing
input data space dimensionality to minimize the computational cost. It
was designed as a natural sub-part of the process of classiication for
many learning algorithms. Generally, three feature selection methods
i.e. ilter method, wrapper method, and embedded method achieve the
objective of selecting important features. The ultimate goal of feature
selection is always to ind the collection of the best features out of the
entire dataset to obtain improved classiication results. Among all of
the feature selection methods, information gain, chi-square, and Gini
index have been used effectively [18], [29], [30]. These methods have
shown promising results for classiication [31]. CHI square relects
one of the more traditional feature selection strategies. In statistics, the
CHI square test is used to analyze the independence of two instances.
The instances, X and Y, are taken as separate if:
(4)
These two instances result in a particular word and class occurring
respectively in the collection of text features. It can be calculated as
given in equation (5):
(5)
Here, N is termed an observed frequency and E is the expected
frequency for every term state t and class C. CHI square would be
the function of how often the expected value E counts and N counts
observed to deviate from one another. A high value of CHI square
means that the independence supposition is wrong. If these two
instances are related, then the term existence increases the probability
of the class existence. This determines the weighted average score
for all classes and then chooses the maximum score between all
classes. In this paper, as in (6) given by [29], the former method is
ideal to globalize the CHI square value for all classes. Here P
(C
i
) is the
likelihood of a class and Chi
2(t, C
i
) is the unique Chi
2
value of a term t.
(6)
Another effective method has been used by researchers i.e.
Information Gain. This assesses the overall knowledge that the
existence or absence of a word allows one to make the right
classiication judgment for every class [32]. In other words, it can be
used in the selection of features by assessing each variable’s gain in
the target variable sense. The measurement between the two random
variables is considered mutual information.
(7)
In equation (7), the total classes are represented by M, probability
of class c is represented by P(c), the presence and absence of term t
are denoted by P(t) and P( ), P(c|t) and P(c| ) are class c possibilities
provided the existence and absence of a term t.
The other ilter method which has been effectively used is the Gini
Index [20]. In general, it has simpler computations than the other
methods. It can be calculated as given in equation (8):
(8)
In (8), P(t/C
i
) is the likelihood of a term t provided that the class C
i
is present. P(C
i
/t) is a class C
i
probability given the presence of term t.
3. Classiication Models
Classiication is a supervised technique of machine learning
wherein the computer algorithm learns from the data it receives as
inputs and then uses the experience to classify new data. This data
collection may be purely binary or multi-class classiication. Types of
classiication tasks include voice recognition, handwriting recognition,
scam detection, news labeling, etc. There has been several machine
learning discovered from time to time with different approach and
application. One of the models is Naive Bayes, simple to build and
use for an extremely large volume of data. The classiier Naive Bayes
claims that every other feature is unrelated to the inclusion of a
speciic feature in a class. Even though these characteristics depend
on each other or the presence of the other characteristics, each of
these properties contributes to the likelihood independently. It can be
calculated as given in equation (9) and (10):
(9)
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 39 -
(10)
Here c refers to class and x represents inputs. Given the data 𝑥,
P(c|𝑥) is mentioned as the posterior probability of c, P(𝑥|c) probability
of input value x provided hypothesis was true, P(c) represents the
prior probability of c, and P(𝑥) is the prior probability of x.
[33] uses Naïve Bayesian classiier along with two feature
evaluation metrics to multi-class text datasets i.e. multi-class Odds
Ratio (MOR) and Class Discriminating Measure (CDM) to achieve the
best feature selecting results. The other k-nearest-neighbors classiier
algorithm takes up a lot of labeled points and uses them to know how
to classify certain items. It looks at the points nearest to the new point
to identify a new point, so whatever label most neighbors have is the
new point label. [16] uses the neighbor-weighted K-nearest neighbor
algorithm achieving signiicant performance gains in the classiication
of an imbalanced data set.
The statistical method, Logistic Regression, is used for evaluating
a data set in which a result is calculated by one or more independent
variables. Uses the probability log-odds of an event that is a linear
combination of independent or prediction variables. Logistic
Regression uses the Sigmoid activation function which results in
either 0 or 1. It can be calculated as given in equation (11):
(11)
Here, z represents the input variable.
It is proven superior to other binary classiication such as KNN, as it
also describes quantitatively the factors leading to classiication [34].
The goal is to identify the best it model to explain the relationship
between the dichotomous value attribute and a series of independent
variables. Decision Tree algorithm gives signiicant results for treating
both categorical and numerical data. In the form of classiication or
regression models, the decision tree builds a tree structure. It splits
down a collection of data into smaller and smaller subsets, thus
constructing a linked decision tree incrementally. The tree splitting
uses Chi-square, Gini-Index, and Information gain methods. A decision
tree with improved chi-square feature selection outperforms in terms
of recall for multiclass text classiication [35].
The various classiiers being studied in the different applications
have shown varied results. The authors have been proposed ensemble
methods to further improve classiication accuracy measures.
Ensemble learning is the mechanism by which several models are
systematically created and merged to solve a speciic computational
intelligence problem. Random forests are an ensemble learning system
for classiication, regression, and other functions that operates by
creating a multitude of decision trees during training and providing
class mode (classiication) or mean forecasting (regression) of the
individual trees. [36] uses ensemble methods for keyword extraction
where Random Forest shows promising results. The authors have been
improving such methods for effective text classiication [37].
In the current scenario where data has been converting to big
data, Neural Networks have been the most studied algorithms for text
classiication. A neural network is a type of layer-organized units
(neurons) that transforms some output into an input vector. Every
unit can take input, impose a function on it, and pass the output to the
next layer. The networks are commonly known as feed-forward: a unit
feeds its output to all the units on the next layer but no input is given
to the previous layer. Weights are added to the signals that travel from
one unit to another, and it is these weights that are adjusted during
the training phase to it a neural network to a speciic problem. [38]
proposes three distinct frameworks for sharing information with task-
speciic and shared layers to model text, based on recurrent neural
networks. These deep learning algorithms’ successes depend on their
ability to model complex and nonlinear interactions within the data.
Finding suitable architectures for these models, however, has been a
problem for researchers addressing leveraging.
B. Techniques for Dealing With Imbalanced Data
We will illustrate in this section the various techniques which have
been experienced so far by researchers for training a model to perform
well against highly imbalanced data sets. The authors mentioned that
where it comes to text classiication, the normal distribution of textual
data is often unbalanced. To better differentiate documents into
minor categories, they used a basic probability-based word weighting
scheme to solve the problem [39]. Many real-world text classiication
tasks, according to the authors, require unbalanced training instances.
However, in the text domain, the methods introduced to resolve
the imbalanced description have not been consistently tested. They
conducted a survey based on the taxonomy of strategies suggested for
imbalanced classiication, such as resampling and instance weighting,
among others [40]. The following sub-sections cover the literature
of various techniques used so far to deal the text classiication with
imbalanced data sets.
1. Data Level Technique
Dealing with imbalanced data sets requires techniques such as
enhancing classiication algorithms or balancing the training data
classes until the machine learning algorithm provides the data as
input. The primary goal of balancing classes is either to raise the
frequency of the minority class or to decrease the frequency of the
majority class. This is provided for all classes to get roughly the same
number of instances.
• Under-sampling aids in optimizing class allocation by randomly
eliminating instances of majority classes. This is achieved when
the majority and minority class cases are balanced completely.
Evolutionary undersampling outperforms the non-evolutionary
models by increasing the degree of imbalance [41]. They
describe a performance function that incorporates two values:
the classiication factor aligned with both the sub-set of training
instances and the percentage of reduction associated with the
training set of the same sub-set of instances. A novel under-
sampling technique is implemented, called cluster-based instance
selection, which incorporates clustering analysis with instance
selection [42]. The clustering analysis framework groups identical
data samples of the majority class dataset into subclasses, while
the instance selection framework extracts out unaccountable data
samples from each subclass. It is also proven that under-sampling
with KNN is the most powerful approach [43].
• Over-sampling raises the amount of minority class instances
by arbitrarily replicating them to make the minority class more
represented in the study. The author suggests a Random Walk Over-
Sampling method by generating synthetic samples by randomly
walking from real data to match different class samples [44]. This
sampling method is designed to address the imbalanced grouping
of data by producing some samples of a synthetic minority class.
The synthetic samples, that properly follow the initial minority
training set and extend the minority class boundaries, are coupled
with the actual samples to make a more eficient full dataset, and
the entire is used to build unbiased classiiers. Unfortunately,
traditional over-sampling approaches have shown their respective
shortcomings, such as causing serious over-generalization or not
effectively improving the class imbalance in data space, while
facing the more challenging problem as opposed to the binary
class imbalance scenario. The author proposes a synthetic minority
oversampling algorithm based on k-nearest neighbors (k-NN),
called SMOM, for handling multi-class imbalance problems [20].
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 40 -
SMOM is a method to prevent over-generalization since safer
neighboring directions are more likely to be chosen to produce
synthetic instances. It is also suggested that combine sampling be
rendered by combining the techniques of SMOTE and Tomek with
SVM as the method of binary classiication [45]. SMOTE is a useful
over-sampling technique for increasing the number of positive
classes incorporating sample drawing methods by replicating the
data randomly so that the number of positive classes is equal to that
of the negative class. [46] has performed multiclass classiication
with equal distribution of the data among various classes using
SMOTE, owing to the introduction of synthetic instances which
increased the number of training samples to distribute the data
equally among 10 different labels. Tomek links method is under-
sampling, which works by decreasing negative class numbers.
However, in some extreme cases mixing sampling methods are no
stronger than utilizing Tomek link methods.
2. Algorithms-Based Decomposition Techniques
The technique must irst use decomposition strategies to transform
the original multi-class data into binary subsets.
• One-vs-all is a strategy that requires training N independent
binary classiiers, each programmed to identify a speciic class.
All those N classiiers are collectively used to classify multiple
classes. With multi-class imbalanced data, an algorithm called
One-vs-All with Data Balancing (OAA-DB) is built to enhance the
classiication performance [47]. It is mentioned that the OAA-DB
algorithm can boost classiication eficiency for imbalanced multi-
class data without decreasing the overall classiication accuracy. In
other words, for every class, One-vs-All trains a single classiier,
treating the existing class as the minority one and the remaining
classes as a majority.
• One-vs-One trains a binary classiier for each potential pair of
classes, ignoring examples that are not part of the pair classes.
To resolve the multi-class imbalance classiication problems,
an exhaustive empirical study is proposed to investigate the
possibility of improving the one-vs-one scheme through the
application of binary ensemble learning approaches [48].
• One-Against-Higher-Order (OAHO) is an explicitly designed
decomposition process for unequaled sets of data. OAHO irst
divides class by decreasing the number of samples [49]. OAHO
sequentially marks the current class as ‘positive class’ and all the
remaining classes with lower ranks as ‘negative classes,’ then
trains a binary classiier.
• All-in-One uses One-vs-All along with One-vs-One, it irst uses
One-vs-All sub-classiiers to ind the top two most probable
categories for each test case, and then use the corresponding One-
Vs-One sub-classiied to decide the inal result [50].
3. Algorithms-Based Ensemble Techniques
The main purpose of the ensemble methodology is to improve
single classiier eficiency. The method involves constructing from
the original data numerous two-stage classiiers and then aggregating
their predictions.
a) Boosting-Based Techniques
One strategy which can be used to increase classiication eficiency
is boosting. Although several data sampling techniques are explicitly
developed to ix the issue of class imbalance, boosting is a technique
that can increase the eficiency of any weak classiier. Ada Boost
iteratively constructs a model ensemble, which is an adaptive boosting
strategy that combines many weak and inaccurate rules to build a
predictive rule that is highly effective. During each iteration, case
weights are changed to properly classify the instances in the next
iteration that were wrongly classiied during the current iteration.
Upon completion, all models developed to take part in a weighted vote
to identify unlabeled cases. Such a strategy is especially useful when
grappling with class inequality as in successive implementations the
minority class instances are more likely to be misclassiied and thus
assigned larger weights. In other words, it’s a binary classiication
algorithm that combines many weak classiiers to create a stronger
classiier [4]. Boosting can be achieved either by “reweighing”
or “resampling”. At each step, the changed example weights are
transferred directly to the base learner while boosting by reweighing.
Not all learning algorithms are designed to integrate example weights
into their decision-making systems, however. This is a class that uses
the AdaBoost Ml method to boost a nominal classiier which can only
address nominal class problems. It is given in equation (12):
(12)
Here, f(𝑥) represents m
th
weak classiier and θ
m
is the corresponding
weight.
This often improves the performance dramatically but sometimes
overits [51]. Gradient boosting is an approach that generates a set
of weak regression trees by introducing iteratively a new one which
further strengthens the learning goal by optimizing an arbitrary
differentiable loss function [52]. Gradient Boosting builds the irst
learner to predict the samples on the training dataset and calculates
the loss. And use that loss in the second stage to build an improved
learner. The recent implementation of this boosting method called
XGBoost combines the principles of computational eficiency. The
paper presents a scalable end-to-end tree boosting system XGBoost
that is widely used by data scientists to perform state-of-the-art
machine learning outcomes [52].
b) Bagging-Based Techniques
Bootstrap aggregation, also known as bagging, is an ensemble meta-
algorithm for machine learning that aims to enhance the stability
and accuracy of classiication algorithms. The standard algorithm
requires the development of speciic bootstrap training items, ‘n’
with substitution. Then train the algorithm on each bootstrapped
algorithm separately, then aggregate the forecasts at the end. The
authors present online bagging and boosting versions that require
only one pass through the training data [53]. Random Forests is an
ensemble classiier composed of several decision trees and generating
the class which is the class output mode for individual trees. In this
way, an RF ensemble classiier works more than a single tree from the
classiication results perspective [54]. The authors suggested ensemble
classiiers focused on original principles such as learning cluster
boundaries by the base classiiers and mapping cluster conidences to
a class decision using a fusion classiication [55]. The classiied data
set is divided into several clusters and is fed into several distinctive
base classiiers. Cluster boundaries are identiied to base classiiers and
cluster conidence vectors are built. A second stage fusion classiier
blends class decisions with conidences and maps of the clusters. This
ensemble classiier restructured the learning environment for the base
classiiers and promoted successful learning.
4. Other Techniques
Despite their effectiveness, however, sampling methods add
complexity and the selection of required parameters. To address these
problems, the author suggests a modern decision tree strategy named
Hellinger Distance Decision Trees (HDDT), which allows the use of
distance from Hellinger as the criteria for splitting. For probability
and statistics, the Hellinger distance is used to measure the correlation
of two distributions of probabilities. The authors use a Hellinger
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 41 -
weighted ensemble of HDDTs to combat deinition drift and improve
the accuracy of single classiiers [56].
Error Correcting Output codes, ECOC is a common multi-class
learning tool that works by breaking down the multi-class task into
a set of binary class subtasks (dichotomies) and creating a binary
classiier from each dichotomy. Both the dichotomy classiiers evaluate
a test instance and then assign it to the nearest class in code space. A
suitable code matrix, an effective learning strategy, and a decoding
strategy highlighting minority classes are needed to enable ECOC
to tackle multi-class imbalances. The authors propose the imECOC
approach that operates on dichotomies to deal with both the imbalance
between class and the imbalance within a class [57]. ImECOC assigns
dichotomy weights and uses weighted decoding distances where
optimum dichotomy weights are derived through reducing weighted
loss in terms of minority classes.
The authors suggest merging weighted One-vs-One voting with
a Winnow dynamic combiner customized to the program for the
data stream. This will allow weights for classiiers to be dynamically
modiied, boosting the power of those competent in the current state
of the stream [17]. DOVO simply adjusts the weights for classiied
objects returned via an active learning approach that enables even
more consistent weights and lower processing costs. From those in the
perspective of operation recognition, each action shall be taken over
a given period. The proposed weighting procedure thereby enables
to rapidly increase the signiicance of qualiied classiiers to identify
this particular behavior immediately after it has been identiied by the
active learning methodology and to sustain the signiicant importance
of these related classiiers throughout its length.
C. Existing Solutions or Software for Classiication With
Imbalanced Datasets
A program, KEEL [58], provides a customized algorithm for the
problem of classiication with class imbalances. Multi-IM draws
its basis from the probabilistic relational methodology (PRMsIM),
developed to learn from imbalanced data for the problem of two
categories [59]. Imbalanced-learn; A Python toolbox for resolving
imbalanced results [60].
We use the following framework to evaluate the accuracy output
of various ML algorithms and to validate our implementations in the
classiication of multi-class imbalance data on inancial news datasets.
III. F W F N
C S: C S
D I
Text classiication is crucial for information extraction and
summarization, text retrieval, and question-answering in general.
Using machine learning algorithms, the authors demonstrated the text
classiication process [19]. Following the approach, we developed a
structure shown in Fig. 2. to distinguish the banking and other related
sector-oriented news items from inancial news posts. It involves three
stages, including the data pre-processing phase, the training phase of
the classiiers, and a comparative estimation of the performance phase
of the classiiers. The phases are discussed in brief in the sub-sections
along with certain challenges and solutions are given by researchers.
However, when faced with imbalanced multi-class results, we can
drop output on one class quickly when attempting to get output on
another class. A clearer analysis of the essence of the issue of class
imbalance is required, as one should recognize in what realms class
imbalance most impedes the output of traditional multi-class classiiers
while developing a system suitable to this topic. Although most of the
problems addressed in the preceded section can be applied to these
multi-class concerns, the banking and other related news extraction
from the inancial news domain. We are identifying the following vital
research directions for the future.
Data pre-processing
steps
Training Classifiers Testing classifiers and
their performance
Data Collection and
cleaning
Selection of classiiers
to be trained on dataset
Selection of
evaluation parameters
Testing the classiiers
on test data and
analyzing their
performance on
dierent metrics
Training the classiiers
on the given data
Manual Labelling
Data transformation
Data reduction
Fig. 2. Multiclass classication of Financial News.
1. Data Pre-processing
Data preprocessing is a method used to transform the raw data
into an effective and functional format. Effective pre-processing of
text data is critical to achieving an appropriate output and better text
classiication quality [61].
Challenge-A: The task of preprocessing data here may be much
more critical than in the case of binary issues. Possible dificulties
can be easily identiied: class overlap can occur in more than two
classes, class label noise can inluence the issue, and class boundaries
may not be speciic. Therefore, effective data cleaning and sampling
techniques must be implemented to take into consideration the
various characteristics of the classes and the balanced performance of
all of them [62].
Solution-1: The problem of noise present in the data in the case
of imbalanced distributions is incredibly dificult. Distortions may
dramatically deteriorate classiier eficiency, particularly in the case
of minority examples. New data cleaning methods need to be used to
manage the existence of overlapping and chaotic samples which can
also lead to worsening eficiency of the classiier. We might conceive
projections into different spaces where the overlap is alleviated or basic
examples are eliminated as mentioned in 3.1.3. However, measures
are needed to assess whether a provided overlapping example can be
excluded without discrimination to one class. A study of the effect on
the real imbalance between classes is quite important in the case of
label noise. Measures are therefore required to determine whether a
given overlapping example can be discarded without compromising
one of the classes. False labeling may lead to increasing the imbalance
or disguise actual disproportions. This situation is handled with
sustained methods for sensing and iltering noise, as well as handling
and relabeling strategies for such examples as mentioned in 1.
Solution-2: Analysis of the kind of examples found in each class
and their connections with other classes is interesting. Measuring
each sample’s dificulty here isn’t straightforward, as it may adjust to
various classes. For instance, for classes Banking and Governmental,
news related to a collective decision on negative GDP outlook and
modiication on repo rate by RBI may be of borderline type while at
the same time being a safe example when considering the remaining
classes. Therefore, we have preferred a more lexible classiication i.e.
SMOTE. SMOTE functions by choosing similar examples in the vector
space, drawing a line through the examples in the vector space, and
drawing a new example at a point in the line.
Solution-3: New sampling approaches are needed for issues of
multiple classes. Simple re-balancing is not a proper approach towards
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 42 -
the largest or smallest class. We need to establish precise methods
for adapting the sampling procedures to both the individual class
property and their mutual relationships. [6] has provided the ensemble
methods to deal with class imbalance classiication, ADASYNBagging,
and RSYNBagging. The ADASYN and RSYN were based on over-
sampling and under-sampling techniques respectively. These were
combined with a bagging algorithm to integrate the advantages of
both algorithms. Another paper has provided a hybrid model to get
a random sample from an unknown population. When compared
with a random sample, a non-random sample could not provide better
representative inferential statistics. Hence, to overcome this problem,
Snoran Sampling Method was developed by [63]. We have not
implemented these techniques in our paper. What sampling strategies
would function best with the learning of the ensemble to boost class
inequality, however, is highly dependent on problem domains.
2. Data Collection
To continue this, we gathered data by scrapping news from public
news sources such as Bloomberg, Financial Express, Money Control,
and Times of India using python-written code. As a result, we have
been collected more than 10000 instances of inancial news articles
from the year 2017 to 2020. The news articles belong to different
sectors or market segments. These are then pre-processed such that
the machine learning algorithms may learn from the training dataset
and adapt them in an acceptable way to the testing data collection.
Therefore, these are pre-processed for the machine learning models
to be explored from the training sample and implemented in an
appropriate format to the test data set.
3. Labeling
The irst step in the pre-processing phase is to label the news from 4
classes to which they belong to the speciic sector. 4-classes are named
as Banking, Global, Governmental, and Non-Banking. We prefer
manual labeling [64] of the news articles with the help of experts of
the inancial domain where overlapping examples were preferred to
discard without damaging one of the classes. Table I mentions the
instance of each class as follows:
T I. S
Source News article Class
Source1
1
The Kolkata-based private sector lender
Bandhan Bank surpassed the market
capitalization of all listed PSU banks except
State Bank of India upon blockbuster stock
market debut on Tuesday after loating
India’s biggest bank IPO earlier this month.
Banking
Source2
2
For India, the current account deicit is
within the comfort zone although it has
widened and the GDP growth is heading
towards 7.5-7.7 percent.
Governmental
Source3
3
The U.S. Federal Reserve has cut its
benchmark interest rate by a half-point-the
biggest reduction, and the irst outside of
scheduled meetings since the 2008 crisis year.
Global
Source4
4
The Nifty50 formed a bearish candle for the
sixth consecutive day in a row and analysts
feel that it will be hard for the index to
breach the 200-DEMA in a hurry.
Non-Banking
1
www.nancialexpress.com
2
www.moneycontrol.com
3
www.bloombergquint.com
4
www.moneycontrol.com
4. Data Cleaning
They are then cleaned because the data can have several sections
that are insigniicant and missing. Data cleaning is done to handle that
portion. It includes absent managing data, noisy data, etc. It helps the
machine learning algorithms to eficiently grasp and operate on them.
5. Data Transformation
The next step, data transformation, is taken to turn the data into
appropriate forms suited to the mining process, and the text of news
articles therein is converted into measures with quantitative values by
constructing a vector set of features. Since data mining is a technique used
for managing enormous quantities of data. In these instances, research
became harder when operating with a huge volume of data. To get rid
of that, we use the strategy of data reduction. This seeks to increase
the capacity of storage and reduce the expense of data collection and
analysis. In other words, in the last step of this stage, the feature vector
is normalized and scaled to prevent an unbalanced dataset.
A. Training Classiiers
Training is the practice of having text that is considered to belong
to the deined classes and creating a classiier based on that known
text. The basic concept is that the classiier accepts a collection of
training data describing established instances of classes and uses the
information obtained from the training data to determine the classes
other unknown content belongs to, by conducting statistical analysis
of training data. We can also use the classiier to derive information
on your data based on the statistical analysis carried out during
the training process. First, we identify the classes on a collection of
training data, and then the classiier uses these classes to evaluate and
decide the classiication of other data. When the classiier assesses
the data, it uses two often contradictory metrics to help decide if the
content found in the new data belongs in or outside a class. Precision,
is the likelihood that what has been labeled as being is actually in that
class. High precision may come at the cost of missing certain results
whose terms match those of other outcomes in other groups. Recall,
the likelihood that an object is listed as being in that class in fact in a
class. High recall may come at the cost of integrating outcomes from
other classes whose terms match those of target class results. We need
to ind the right balance with high precision and high recall while we
are tuning our classiier. The balance focuses on what our priorities
and criteria are for implementation. We need to train the classiier
with sample data that describes members of all the classes to ind the
best thresholds for our data. Finding good training samples is very
critical because the nature of the training can directly inluence the
quality of the classiication. The samples should be statistically valid
for each class and should include samples that include both solid class
examples and samples near the class boundary.
Challenge-B: The strong potential resides in the complexity of
multi-class, distort-insensitive classiiers. They will permit multi-
class complications to be handled without referring to strategies for
resampling when algorithm-level approaches are used to counter class
imbalances. So one may wonder if other prominent classiiers can be
adapted in this case [62].
Solution-1: Certain issues should be addressed when constructing
multi-level classiiers in the case of class imbalances. A broader study
is required of how numerous unbalanced data sets inluence decision
boundaries in classiiers. Based on [65] Hellinger distance has proved
useful in cases of class imbalance. Since accuracy may offer a distorted
picture of success on unbalanced data, current stream classiiers are
focused on accuracy that is hampered by minority class output on
unbalanced streams, resulting in low recall levels of minority classes.
A split based on Hellinger Distance will give a high score to a split
separating the classes in the best way relative to the parent population.
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 43 -
When utilizing Hellinger, it is possible to obtain a statistically relevant
change in the recall level on imbalanced data sources, with a reasonable
rise in the false positive rate.
Solution-2: Other solutions with potential robustness to the
imbalance, such as methods based on density, need to be explored. [66]
have provided a more thorough review of the cluster oversampling
based on density and in terms of density-dependent clustering under-
sampling techniques. Their indings suggest the strategy will boost the
classiier’s predictive eficiency. It also yields the best in the precision
average.
Solution-3: While modern methods of learning with imbalances are
suggested to tackle the question of data imbalances, they have certain
limitations; under-sampling methods lose essential details, and cost-
sensitive methods are prone to outliers and noise. [67] has provided an
eficient hybrid ensemble classiier architecture incorporating density-
based under-sampling and cost-effective approaches by investigating
state-of-the-art solutions using an algorithm for multi-objective
optimization. First, they developed a density-based under-sampling
method to select informative samples with probability-based data
transformation from the original training data, which enables multiple
subsets to be obtained following a balanced class-wide distribution.
Second, they have used the cost-sensitive approach of classiication
to address the problem of information incompleteness by modifying
weights in minority groups misclassiied, rather than majority samples.
Finally, they implemented a multi-objective optimization method and
used sample-to-sample relations to auto-modify the classiication
outcome utilizing an ensemble classiication system [68-80].
B. Testing Classiiers and Their Performance
We run the trained classiier on unknown news articles to check
a classiier to decide which classes each news article belongs to. The
goal of this stage is to check the performance of the classiiers on
the training set and to see if they detect the training correctly. The
classiiers considered will be graded according to their effectiveness in
detecting the appropriate class. In the later section, we will test various
classiiers on the unseen news articles and compare the performance
of each.
IV. W F N C S
Throughout this section, we describe irst the experimental method
used to train the classiiers and then demonstrate their success in the
classiication of news articles into four separate classes. It should be
noted here that most text classiication algorithms are prone to the
form and design of the dataset, depending on factors such as class
size, class disparity (number of samples per class), feature scaling,
number of training samples, number of features, etc. Besides, different
algorithms follow different approaches to solving problems of multi-
class classiication which also affects their performance. So, we have
faced some challenges and, to address these challenges, we have made
sure that the available data from which each classiier will learn is
distributed equally for each class.
A. Experimental Set Up to Train the Classiiers
We used the Tableau prep tool for the data cleaning and
preprocessing operations, while the desktop tool was used for the data
visualization. The classiication tests were performed on Python 3.8
utilizing numerous Python-supported libraries to incorporate machine
learning and deep learning algorithms. With a split of 75% and 25%
respectively, the total of 10,000 news articles is divided into training
and test data. The news articles are related to 4 different classes as
mentioned in the introductory section. The data was imbalanced. So,
to balance the data various sampling techniques were used. As stated
in the introductory section, the news articles are linked to 4 different
classes. In nature, the data had been imbalanced. Therefore, different
sampling strategies were used to balance the data among classes
as discussed in section 4.2. The machine and deep algorithms were
further implemented on data for classiication using scikit-learn and
imblearn libraries of Python. Scikit-learn offers a package named the
TidVectorizer for the extraction of functionality from text documents.
This class is responsible for both vectorizing text documents (news
articles) into vectors of word features and transforming them of the
term vectors in the scores of TfIdf. We also vectorized the dataset
during the experiments using the N-gram approach, with unigrams,
bigrams, and tri-grams.
B. Results and Discussion
We have carried out several experiments on our pre-processed data
collection utilizing conventional machine learning algorithms detailed
in the section preceding. The key purpose of these experiments is to
determine the right classiier that gives the best performance. Every
classiier’s output concerning classiication is calculated using the
metrics Precision, Recall, and F
1
-score. The accuracies are obtained
with both train-test split and 5-fold cross-validation for all classiiers.
The outcomes of the chosen classiiers are described in the sub-
sections that follow.
For the traditional machine learning algorithms, TF-IDF features of
1-gram, 2-gram, and 3-gram were used. The detailed experiments on
the inancial news datasets were carried out.
1. Results of Multiclass Classiication With Data-Imbalances
Table II lists the results of each of the classiiers where data
is being vectorized using the N-gram TF-IDF feature with data
imbalances across classes. From the different classiiers Decision
Tree {criterion=’gini’ to measure qualility of split, splitter=’best’,
max_depth=2 for maximum depth of tree, random_state=1 is the
seed for random number generator}, Linear SVC {C=1 regularization
parameter, multi_class=‘y’}, Logistic Regression {C=1 regularization
parameter, random_state=0}, Multinomial Naïve Bayes {alpha=1.0
smoothing parameter}, Random Forest {n_estmators=100 for number
of trees, random_state=1 will always produce same results with same
parameters and training data, max_depth=3 for maximum depth
of the tree}, and Multilayer Perceptron {solver=’lbfgs’ for weight
optimization, alpha=0.0001 L2 penality, learning_rate=’constant’,
hidden_layer_sizes=(5,2), random_state=1}, Random Forest performed
best with accuracy 88% as shown in Table III. The Random Forest
TABLE II. R C D C W D I
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
Decision Tree 1, 2, 3 0.93 0.96 0.94 0.76 0.68 0.72 0.90 0.95 0.92 0.80 0.31 0.44
Linear SVC 1, 2, 3 1.00 0.77 0.87 0.86 0.61 0.71 0.86 0.98 0.91 1.00 0.38 0.56
Logistic Regression 1, 2, 3 1.00 0.31 0.47 0.88 0.34 0.49 0.76 0.99 0.86 0.00 0.00 0.00
Multinomial NB 1, 2, 3 0.86 0.23 0.36 0.73 0.54 0.62 0.79 0.97 0.87 0.00 0.00 0.00
Random Forest 1, 2, 3 0.93 1.00 0.96 0.89 0.59 0.71 0.88 0.98 0.92 1.00 0.23 0.38
Multi-layer Perceptron 1, 2, 3 1.00 0.69 0.82 0.78 0.68 0.73 0.86 0.96 0.91 1.00 0.31 0.47
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 44 -
achieved the F
1
-score 0.96, 0.71, 0.92, 0.38 for classes Banking, Global,
Non-Banking, and Governmental respectively. The comparison of all
the mentioned classiiers for 4-different classes is visualized in Fig. 3-6.
Table III shows that the accuracy comes out to be 78%-88% range for all
machine learning algorithms with train-test split and cross-validation.
TABLE III. A C W I D
Classifier Accuracy(Train/Test) Cross-Validation
Decision Tree 0.87 0.87
Linear SVC 0.87 0.87
Logistic Regression 0.78 0.78
Multinomial NB 0.78 0.78
Random Forest 0.88 0.88
Multi-layer Perceptron 0.87 0.87
However, Table II shows that the recall of the minority classes is
very less. The Logistic Regression and Multinomial NB has shown
0% precision and recall for the minority class i.e. Governmental. This
is visualized in Fig. 5. At the same time, the precision and recall for
the other classes have shown high precision and recall. It shows that
machine learning models are more biased towards the majority class.
So, we need to apply imbalanced data handling techniques.
2. Results of Multiclass Classiication With Data Balance
Using Data-Level Technique: Random Over-Sampling With
Replacement
The Resampling takes place with the exclusion of the minority
class, increasing the sample number to equal that of the majority class.
Tables IV and V lists the results of each of the classiiers where data is
being vectorized using the N-gram TF-IDF feature with data balanced
using random over-sampling technique across classes.
From the different classiiers Decision Tree, Linear SVC, Logistic
Regression, Multinomial Naïve Bayes, Random Forest, and Multilayer
Perceptron with accuracy for all classes with balanced datasets using
up-sampling, the Random Forest again performed best with accuracy
99% as shown in Table V. The Random Forest achieved the F
1
-score
1.00, 0.98, 0.98, 1.00 for classes Banking, Global, Non-Banking, and
Governmental respectively. The comparison of all the mentioned
classiiers for 4-different classes is shown in Tables IV and V. It is
observed that with data balances the precision and recall have also
Classifier
Decision Tree
Decision
Tree
Linear SVC
Linear
SVC
Logistic Regression
Logistic
Regression
Multi-layer Perceptron
Multi-layer
Perceptron
Multinomial NB
Multinomial
NB
Random Forest
Random
Forest
0.0
0.9200
0.9100
0.8600
0.9100
0.8700
0.9200
0.9500
0.9800
0.9900
0.9600
0.9700
0.9800
0.9000
0.8600
0.7600
0.8600
0.7900
0.8800
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
F1 R P
Fig. 5. Performance metrics P, R, F-1 for various classiers for Non-Banking
Class.
Classifier
Decision Tree
Decision
Tree
Linear SVC
Linear
SVC
Logistic Regression
Logistic
Regression
Multi-layer Perceptron
Multi-layer
Perceptron
Multinomial NB
Multinomial
NB
Random Forest
Random
Forest
0.0
0.9600 0.9400
0.8700
0.8200
0.4700
0.3400
1.0000
0.9600
0.7700
0.6900
0.3100
0.2300
0.9300 0.9300
1.0000 1.0000 1.0000
0.8600
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
F1 R P
Fig. 3. Performance metrics P, R, F-1 for various classiers for Banking Class.
Classifier
Decision Tree
Decision
Tree
Linear SVC
Linear
SVC
Logistic Regression
Logistic
Regression
Multi-layer Perceptron
Multi-layer
Perceptron
Multinomial NB
Multinomial
NB
Random Forest
Random
Forest
0.0
0.7200 0.7100
0.4900
0.7300
0.6200
0.7100
0.6800
0.6100
0.3400
0.6800
0.5400
0.5900
0.7600
0.8600
0.8800
0.7800
0.7300
0.8900
0.4
0.6
0.2
0.8
0.0
0.4
0.6
0.2
0.8
0.0
0.5
1.0
F1 R P
Fig. 4. Performance metrics P, R, F-1 for various classiers for Global Class.
Classifier
Decision Tree
Decision
Tree
Linear SVC
Linear
SVC
Logistic Regression
Logistic
Regression
Multi-layer Perceptron
Multi-layer
Perceptron
Multinomial NB
Multinomial
NB
Random Forest
Random
Forest
0.0
0.4400
0.5600
0.0000
0.4700
0.3800
0.3100
0.3800
0.0000
0.3100
0.2300
0.8000
1.000
0.0000
0.0000
0.0000
0.0000
1.000 1.000
0.2
0.4
0.6
0.0
0.2
0.4
0.0
0.5
1.0
F1 R P
Fig. 6. Performance metrics P, R, F-1 for various classiers for Governmental
Class.
TABLE IV. R C D C W B D U US
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
Decision Tree 1, 2, 3 0.99 1.00 0.99 0.95 0.99 0.97 0.99 0.94 0.97 0.99 1.00 1.00
Linear SVC 1, 2, 3 0.98 1.00 0.99 0.98 0.99 0.98 0.99 0.97 0.98 0.99 1.00 1.00
Logistic Regression 1, 2, 3 0.98 1.00 0.99 0.93 1.00 0.96 1.00 0.91 0.95 0.99 1.00 1.00
Multinomial NB 1, 2, 3 0.97 0.96 0.97 0.93 0.97 0.95 0.93 0.83 0.88 0.94 1.00 0.97
Random Forest 1, 2, 3 0.99 1.00 1.00 0.98 0.99 0.98 0.99 0.98 0.98 1.00 1.00 1.00
Multi-layer Perceptron 1, 2, 3 0.99 1.00 0.99 0.94 0.99 0.97 0.99 0.93 0.96 1.00 1.00 1.00
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 45 -
improved for every classiier. And the accuracy of the classiiers varies
between 94% to 100% and it is visualized in Fig. 7.
TABLE V. A C W B D U U
S
Classifier Accuracy(Train/Test) Cross-Validation
Decision Tree 0.98 0.983
Linear SVC 0.98 0.982
Logistic Regression 0.98 0.975
Multinomial NB 0.94 0.948
Random Forest 0.99 0.996
Multi-layer Perceptron 0.98 0.985
Classifier1
Decision Tree
Linear SVC
Logistic Regression
Multi-layer Perceptron
Multinomial NB
Random Forest
0.0
0.2
0.4
0.6
Accuracy
0.8
1.0
Decision
Tree
0.9800
Linear
SVC
0.9800
Logistic
Regression
0.9800
Multinomial
NB
0.9400
Multi-layer
Perceptron
0.9800
Random
Forest
1.000
Fig. 7. Accuracy for various classiers with balanced classes using Up-
Sampling.
3. Results of Multiclass Classification With Data Balance
Using Data-Level Technique: Random Down-Sampling Without
Replacement
This is done by resampling the majority class without replacement,
setting the number of samples corresponding to that of the minority
class. Table VI, VII lists the results of each of the classiiers where
data is being vectorized using the N-gram TF-IDF feature with data
balanced using down-sampling technique across classes.
From the different classiiers Decision Tree, Linear SVC, Logistic
Regression, Multinomial Naïve Bayes, Random Forest, and Multilayer
Perceptron with accuracy for all classes with balanced datasets using
down-sampling, the Random Forest again performed best with an
accuracy of 80% as shown in Table VII.
The Random Forest achieved the F
1
-score 0.95, 0.83, 0.73, 0.70 for
classes Banking, Global, Non-Banking, and Governmental respectively.
The comparison of all the mentioned classiiers for 4-different classes
is shown in Tables VI and VII. The accuracy of the classiiers has
degraded with data balances using down-sampling as compared to up-
sampling. And the accuracy of the classiiers varies between 67% to
80% and it is visualized in Fig. 8.
TABLE VII. A C W B D U
DS
Classifier Accuracy(Train/Test) Cross-Validation
Decision Tree 0.69 0.695
Linear SVC 0.76 0.764
Logistic Regression 0.71 0.720
Multinomial NB 0.67 0.750
Random Forest 0.80 0.803
Multi-layer Perceptron 0.69 0.692
Classifier1
Decision Tree
Linear SVC
Logistic Regression
Multi-layer Perceptron
Multinomial NB
Random Forest
0.0
0.2
0.3
0.1
0.4
0.6
0.5
Accuracy
0.8
0.7
Decision
Tree
0.6900
Linear
SVC
0.7600
Logistic
Regression
0.7100
Multinomial
NB
0.6700
Multi-layer
Perceptron
0.6900
Random
Forest
0.8000
Fig. 8. Accuracy for various classiers with balanced classes using Down-
Sampling.
4. Results of Multiclass Classiication With Data Balance Using
Data-Level Technique: Hybrid Over-Sampling Technique SMOTE
SMOTE helps to balance the representation of the classes by
replicating randomly through minority class examples. SMOTE
synthesizes new instances within existing instances of minority classes.
This produces the virtual train records by linear interpolation for the
minority class. For each case, these synthetic training records are
created by a random selection of one or more k-nearest neighbors in the
minority class. The data is reconstructed after the oversampling process,
and the classiication models are implemented for the processing data.
Table VIII lists the results of each of the classiiers where data is being
vectorized using the N-gram TF-IDF feature with data balanced using
the over-sampling technique SMOTE across classes.
From the different classiiers Decision Tree, Linear SVC, Logistic
Regression, Multinomial Naïve Bayes, Random Forest, and Multilayer
Perceptron with accuracy for all classes with balanced datasets using
up-sampling, the Random Forest again performed best with accuracy
100% as shown in Table IX. The Random Forest achieved the F
1
-score
0.99, 1.00, 0.99, 1.00 for classes Banking, Global, Non-Banking, and
Governmental respectively. The comparison of all the mentioned
classiiers for 4-different classes is shown in Tables VIII and IX. It
is observed that with data balances using SMOTE the precision and
recall have also improved for every classiier. And the accuracy of the
classiiers varies between 94% to 100% as visualized in Fig. 9.
TABLE VI. R C D C W B D U DS
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
Decision Tree 1, 2, 3 1.00 0.91 0.95 0.56 0.83 0.67 0.71 0.38 0.50 0.60 0.67 0.63
Linear SVC 1, 2, 3 1.00 0.91 0.95 0.75 0.80 0.77 0.78 0.67 0.72 0.67 0.67 0.67
Logistic Regression 1, 2, 3 1.00 0.82 0.90 0.69 0.75 0.72 0.70 0.54 0.61 0.54 0.78 0.64
Multinomial NB 1, 2, 3 0.82 0.82 0.82 0.69 0.75 0.72 0.67 0.31 0.42 0.53 0.89 0.67
Random Forest 1, 2, 3 1.00 0.91 0.95 0.83 0.83 0.83 0.89 0.62 0.73 0.57 0.89 0.70
Multi-layer Perceptron 1, 2, 3 0.88 0.64 0.74 0.71 0.83 0.77 0.62 0.62 0.62 0.60 0.67 0.63
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 46 -
TABLE IX. A C W B D U SMOTE
US
Classifier Accuracy(Train/Test) Cross-Validation
Decision Tree 0.98 0.981
Linear SVC 0.98 0.948
Logistic Regression 0.98 0.972
Multinomial NB 0.94 0.948
Random Forest 1.00 0.995
Multi-layer Perceptron 0.98 0.986
Classifier1
Decision Tree
Linear SVC
Logistic Regression
Multi-layer Perceptron
Multinomial NB
Random Forest
0.0
0.2
0.4
0.6
Accuracy
0.8
1.0
Decision
Tree
Linear
SVC
Logistic
Regression
Multinomial
NB
0.94000
Multi-layer
Perceptron
0.980000.980000.980000.98000
Random
Forest
1.0000
Fig. 9. Accuracy for various classiers with balanced classes using SMOTE
Up-Sampling.
5. Results of Multiclass Classiication With Data Balance Using
Data-Level Technique: Over-Sampling Technique ADASYN
ADASYN (Adaptive synthetic sampling approach) algorithm builds
on the methodology of SMOTE. This uses a weighted distribution for
speciic examples of minority classes due to their degree of learning
capacity, whereas more sophisticated data is generated for examples
of minority classes that are more dificult to understand. The key
idea of the ADASYN algorithm is to use a density distribution as a
parameter to automatically calculate the number of synthetic samples
that each minority data example requires to be generated. The data is
reconstructed after the oversampling process, and the classiication
models are implemented for the processing data. Table X lists the
results of each of the classiiers where data is being vectorized using
the N-gram TF-IDF feature with data balanced using the over-sampling
technique ADASYN across classes.
From the different classiiers Decision Tree, Linear SVC, Logistic
Regression, Multinomial Naïve Bayes, Random Forest, and Multilayer
Perceptron with accuracy for all classes with balanced datasets using
up-sampling, the Random Forest again performed best with accuracy
91% as shown in Table XI. The Random Forest achieved the F
1
-score
0.94, 0.79, 0.94, 0.53 for classes Banking, Global, Non-Banking, and
Governmental respectively. The comparison of all the mentioned
classiiers for 4-different classes is shown in Tables X and XI. It is
observed that with data balances using ADASYN the precision and
recall have also downgraded as compared to SMOTE based up-
sampling for every classiier. And the accuracy of the classiiers varies
between 72% to 91% as visualized in Fig. 10.
TABLE XI. A C W B D U
ADASYN US
Classifier Accuracy(Train/Test) Cross-Validation
Decision Tree 0.87 0.872
Linear SVC 0.87 0.865
Logistic Regression 0.88 0.881
Multinomial NB 0.72 0.725
Random Forest 0.91 0.914
Multi-layer Perceptron 0.86 0.863
Classifier1
Decision Tree
Linear SVC
Logistic Regression
Multi-layer Perceptron
Multinomial NB
Random Forest
0.0
0.2
0.3
0.1
0.4
0.6
0.5
Accuracy
0.8
0.9
0.7
Decision
Tree
0.8700
Linear
SVC
0.8700
Logistic
Regression
0.8800
Multinomial
NB
0.7200
Multi-layer
Perceptron
0.8600
Random
Forest
0.9100
Fig. 10. Accuracy for various classiers with balanced classes using ADASYN
Up-Sampling.
TABLE VIII. R C D C W B D U SMOTE
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
Decision Tree 1, 2, 3 0.99 0.93 0.96 0.99 1.00 1.00 0.95 0.99 0.97 0.98 1.00 0.99
Linear SVC 1, 2, 3 0.99 0.92 0.96 0.98 1.00 0.99 0.94 0.99 0.97 0.99 1.00 1.00
Logistic Regression 1, 2, 3 1.00 0.91 0.95 0.98 1.00 0.99 0.93 1.00 0.96 0.99 1.00 1.00
Multinomial NB 1, 2, 3 0.92 0.85 0.88 0.97 0.96 0.97 0.94 0.96 0.95 0.94 1.00 0.97
Random Forest 1, 2, 3 0.99 0.99 0.99 0.99 1.00 1.00 0.99 0.99 0.99 1.00 1.00 1.00
Multi-layer Perceptron 1, 2, 3 0.99 0.93 0.96 0.99 1.00 0.99 0.94 0.99 0.97 1.00 1.00 1.00
TABLE X. R C D C W B D U ADASYN
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
Decision Tree 1, 2, 3 0.89 0.96 0.93 0.82 0.66 0.73 0.91 0.94 0.92 0.42 0.38 0.40
Linear SVC 1, 2, 3 0.96 0.85 0.90 0.76 0.71 0.73 0.89 0.95 0.92 0.62 0.38 0.48
Logistic Regression 1, 2, 3 0.96 0.85 0.90 0.79 0.76 0.73 0.90 0.95 0.92 0.60 0.46 0.52
Multinomial NB 1, 2, 3 0.50 0.85 0.63 0.57 0.93 0.70 0.99 0.65 0.78 0.29 0.77 0.43
Random Forest 1, 2, 3 0.93 0.96 0.94 0.91 0.71 0.79 0.91 0.98 0.94 0.88 0.38 0.53
Multi-layer Perceptron 1, 2, 3 0.94 0.65 0.77 0.73 0.73 0.73 0.88 0.95 0.92 0.88 0.38 0.53
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 47 -
6. Results of Multiclass Classiication With Data Balances Using
Data-Level Technique: Down-sampling Technique Near-Miss
The NearMiss Algorithm under-sampled the majority class’s
instances and made them equivalent to the minority class. The majority
classes, here, were reduced to the minimum number as of minority
class so that all classes would have the same number of records. The
data is reconstructed after the down-sampling process using the Near-
Miss method, and the classiication models are implemented for the
processing data. Table XII lists the results of each of the classiiers
where data is being vectorized using the N-gram TF-IDF feature with
data balanced using the down-sampling technique Near-Miss across
classes.
From the different classiiers Decision Tree, Linear SVC, Logistic
Regression, Multinomial Naïve Bayes, Random Forest, and Multilayer
Perceptron with accuracy for all classes with balanced datasets using
down-sampling, the Random Forest again performed best with an
accuracy of 81% as shown in Table XIII. The Random Forest achieved
the F
1
-score 0.83, 0.30, 0.89, 0.52 for classes Banking, Global, Non-
Banking, and Governmental respectively. The comparison of all the
mentioned classiiers for 4-different classes is shown in Table XII and
XIII. The accuracy of the classiiers has degraded with data balances
using down-sampling with the Near-Miss approach as compared to all
other up-sampling approaches as visualized in Fig. 11.
TABLE XIII. A C W B D U
NM DS
Classifier Accuracy(Train/Test) Cross-Validation
Decision Tree 0.65 0.652
Linear SVC 0.53 0.526
Logistic Regression 0.34 0.344
Multinomial NB 0.43 0.431
Random Forest 0.81 0.814
Multi-layer Perceptron 0.70 0.704
Classifier1
Decision Tree
Linear SVC
Logistic Regression
Multi-layer Perceptron
Multinomial NB
Random Forest
0.0
0.2
0.3
0.1
0.4
0.6
0.5
Accuracy
0.8
0.7
Decision
Tree
0.6500
Linear
SVC
0.5300
Logistic
Regression
0.3400
Multinomial
NB
0.4300
Multi-layer
Perceptron
0.7000
Random
Forest
0.8100
Fig. 11. Accuracy for various classiers with balanced classes using Near-Miss
Down-Sampling.
7. Results of Multiclass Classiication With Data Balance Using
Ensemble Classiiers
Ensemble models are meta-algorithms incorporating many strategies
in machine learning into one predictive model to minimize variance
(bagging), bias (boosting), or strengthen predictions (stacking). Bagging
methods build multiple estimators on various randomly chosen subsets
of data in ensemble classiiers. The classiier is called BaggingClassiier
in scikit-learn. This classiier, however, does not require a balancing of
the data sub-set. So, this classiier would support the plurality groups
when training on imbalanced data set. BalancedBaggingClassifier
requires each subset of data to be resampled until any of the
ensemble estimators are equipped. In brief, the performance of an
EasyEnsemble sampler is paired with an ensemble of classiiers (i.e.,
BaggingClassiier). Hence the BalancedBaggingClassiier requires the
same parameters as the BaggingClassiier scikit-learn. Additionally,
there are two additional parameters to monitor the actions of
the random under-sampler, sampling strategy, and substitution.
BalancedRandomForestClassifier is another ensemble method
that provides a balanced bootstrap sample for each tree in the forest.
RUSBoostClassifier sub-sample the data collection randomly before
executing a boosting iteration. A particular method in the bagging
classiier which uses AdaBoost as learners is named EasyEnsemble.
The EasyEnsembleClassifier allows AdaBoost learners to be trained
on appropriate samples of bootstrap. Table XIV lists the results of each
of these ensemble classiiers for the various classes. And Table XV
shows the accuracy of these ensemble classiiers.
From the different ensemble classiiers BalancedBaggingClassiier,
BalancedRandomForestClassiier, RUSBoostClassiier, EasyEnsemble
Classiier with accuracy for all classes, the BalancedBaggingClassiier
performed best with an accuracy of 99% as shown in Table XV. The
BalancedBaggingClassiier achieved the F
1
-score 0.97, 1.00, 0.98,
1.00 for classes Banking, Global, Non-Banking and Governmental
respectively. The comparison of all the mentioned ensemble classiiers
for 4-different classes is shown in Table XIV and XV. The accuracy of the
BalancedBaggingClassiier has resulted in 99% which is quite similar
to the result of multiclassiication using Random-Forest Classiier
with SMOTE sampling i.e. 100%. The accuracy of the Random Forest
classiier with a random up-sampling approach for data balances is
also 99%. The comparison of the accuracies of classiiers across all
approaches is visualized in Fig. 12. The accuracy of classiiers with
down-sampling using the Near-Miss approach is worst amongst all.
The accuracy, precision, recall, and F-1 of Random-Forest Classiier
with SMOTE sampling is very good in terms of multiclass news
classiication. However, under Governmental and Banking classes
(minor classes in original), the precision of Random Forest with SOMTE
overlapped with the precision of Random Forest with a random up-
sampling approach. The comparison of the Precision of classiiers
with each approach across all mentioned classes is visualized in Fig.
13. Some of the key explanations for the low performance of some of
the classiiers, including Linear SVC and Multinomial naïve Bayes, is
that a huge number of features don’t it well for them. Earlier it has
been stated that Multinomial Naive Bayes’ output is very weak when
TABLE XII. R C D C W B D U NM
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
Decision Tree 1, 2, 3 0.68 1.00 0.81 0.23 0.24 0.24 0.82 0.70 0.75 0.25 0.54 0.34
Linear SVC 1, 2, 3 0.41 0.96 0.57 0.30 0.56 0.39 0.88 0.45 0.59 0.27 0.69 0.39
Logistic Regression 1, 2, 3 0.43 0.96 0.60 0.29 0.56 0.38 0.94 0.17 0.28 0.12 0.92 0.22
Multinomial NB 1, 2, 3 0.32 0.88 0.47 0.30 0.59 0.40 0.91 0.32 0.47 0.20 0.77 0.32
Random Forest 1, 2, 3 0.91 0.77 0.83 0.67 0.20 0.30 0.82 0.97 0.89 0.60 0.46 0.52
Multi-layer Perceptron 1, 2, 3 0.46 0.81 0.58 0.57 0.61 0.59 0.91 0.72 0.80 0.28 0.62 0.38
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 48 -
the dataset faces class imbalance problems. The result has shown that
the eficiency of the RUSBoostClaasifer ensemble algorithm is very
poor when it comes to the multi-class classiication of text with noisy
data and class imbalance.
TABLE XV. A E C
Classifier
Accuracy
(Train/Test)
Cross-
Validation
BalancedBaggingClassiier 0.99 0.991
BalancedRandomForestClassiier 0.82 0.823
RUSBoostClassiier 0.34 0.344
EasyEnsembleClassiier 0.78 0.781
Approach
ADASYN_UpSampling
Imbalanced
NearMiss_DownSampling
Ramdom_DownSampling
0.0
0.2
0.1
0.3
0.4
0.5
0.6
0.7
Accuracy
0.8
1.0
0.9
Ramdom_UpSampling
SMOTE_UpSampling
Decision
Tree
Linear
SVC
Logistic
Regression
Multinomial
NB
Multi-layer
Perceptron
Random
Forest
Fig. 12. Comparison of accuracies with Classiers across dierent approaches.
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
Non-BankingGovernmental Global Banking
Approach
ADASYN_UpSampling
Imbalanced
NearMiss_DownSampling
Ramdom_DownSampling
Ramdom_UpSampling
SMOTE_UpSampling
Multinomial
NB
Linear
SVC
Classifier
Logistic
Regression
Decision
Tree
Multi-layer
Perceptron
Random
Forest
Fig. 13 Comparison of Precision with Classiers under each class across
dierent approaches.
It is clear from the Fig. 14., the recall of the classiier Random
Forest with data balanced across classes using random up-sampling
and SMOTE is increased as compared to down-sampling techniques
random down-sampling and Near-Miss. The comparison of recall
across all approaches with different classiiers under each class is
visualized in Fig. 14.
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
GovernmentalNon-Banking Global Banking
Approach
ADASYN_UpSampling
Imbalanced
NearMiss_DownSampling
Ramdom_DownSampling
Ramdom_UpSampling
SMOTE_UpSampling
Logistic
Regression
Decision
Tree
Multi-layer
Perceptron
Random
Forest
Multinomial
NB
Linear
SVC
Fig. 14. Comparison of Recall with Classiers under each class across dierent
approaches.
Classifier1
BalancedBaggingClassifier
BalancedRandomForestClassifier
EasyEnsembleClassifier
Random Forest
RUSBoostClassifier
0.0
0.2
0.4
0.6
Accuracy
0.8
1.0
RUSBoost
Classifier
0.3400
EasyEnsemble
Classifier
0.7800
Balanced
Random...
0.8200
Random
Forest
1.0000
Balanced
Bagging...
0.9900
Fig. 15. Ensemble classiers vs Random Forest (SMOTE).
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
GovernmentalNon-Banking Global Banking
Classifier
BalancedBaggingClassifier
BalancedRandomForestClassifier
EasyEnsembleClassifier
RamdomForest
RUSBoostClassifier
EasyEnsemble
Classifier
RUSBoost
Classifier
Balanced
Bagging...
Random
Forest
Balanced
Random...
Fig. 16. Precision of ensemble classiers vs Random Forest (SMOTE).
TABLE XIV. R E C D C
Classifier N-Gram
Banking Global Non-Banking Governmental
P R F
1
P R F
1
P R F
1
P R F
1
BalancedBaggingClassiier 1, 2, 3 0.98 0.97 0.97 0.99 1.00 1.00 0.98 0.98 0.98 0.99 1.00 1.00
BalancedRandomForestClzssiier 1, 2, 3 0.86 0.64 0.73 0.99 0.90 0.94 0.66 0.88 0.76 0.84 0.87 0.86
RUSBoostClassiier 1, 2, 3 0.33 1.00 0.50 0.94 0.34 0.50 0.08 0.05 0.06 0.00 0.00 0.00
EasyEnsembleClassiier 1, 2, 3 0.87 0.93 0.90 0.57 0.44 0.49 0.26 0.85 0.40 0.92 0.83 0.87
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 49 -
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
0.0
0.5
1.0
GovernmentalNon-Banking Global Banking
Classifier1
BalancedBaggingClassifier
BalancedRandomForestClassifier
EasyEnsembleClassifier
RamdomForest
RUSBoostClassifier
EasyEnsemble
Classifier
RUSBoost
Classifier
Random
Forest
Balanced
Random...
Balanced
Bagging...
Fig. 17. Recall of ensemble classiers vs Random Forest (SMOTE).
The accuracy of the ensemble classiiers is compared with Random
Forest with SMOTE and it is visualized in Fig. 15. The accuracy of
multi-class inancial news classiication using Random Forest with
data balanced using SMOTE is higher as compared to all other
ensemble classiiers discussed in the previous section. It is slightly
greater than BalancedBaggingClassiier. The precision and recall of
Random Forest with data balanced using SMOTE across all classes are
higher as compared to all other ensemble classiiers and it is visualized
in Fig. 16. and 17 respectively.
V. C F D
This paper aims to extract banking news from the pool of articles
on inancial news. This multi-class Financial News classiication will
help to get news on the banking domain. The development of a system
for gathering banking news and other relevant domains is a major
and untested problem for the Indian stock market. We’re interested
in seeking news from Indian banks, the Indian government, and the
global. We take a structured approach to divide the news into realms
of our choosing, grouping the news articles into 4 classes. The news
articles are gathered from numerous online news sources and labeled
to derive the banking and other related news to achieve the paper’s
goal. To automate the classiication process, 5 traditional machine
learning classiiers, 1 neural network classiier, and 4 ensemble
classiiers are used to classify the news articles into 4 classes (Banking,
Governmental, Global, and Non-Banking). Since our data set faces
the class imbalance issue, we used many methods to align the data
set between classes, and the classiier output is evaluated using the
original imbalanced and balanced data set. We used precision, recall,
F-1, and accuracy parameters to evaluate the classiication models. It
is evident from results that Random Forest with balanced data using
SMOTE achieved the highest accuracy of 100% whereas other models
have lower classiication accuracy even with 34%. Based on our results,
our trained classiication model can be used to classify the news into
other speciic domains by training the model on data-sets of those
domains. The labeling of the dataset is done manually at the current
stage of our study, with the help of the domain experts. In our future
research, including those listed in this paper, we may also use certain
recently introduced methods and frameworks for classifying data with
a larger volume.
R
[1] Atkins, Adam, Mahesan Niranjan, and Enrico Gerding. “Financial news
predicts stock market volatility better than close price,”The Journal of
Finance and Data Science4, no. 2, pp. 120-137, 2018.
[2] Belainine, Billal, Alexsandro Fonseca, and Fatiha Sadat. “Named entity
recognition and hashtag decomposition to improve the classiication of
tweets,” InProceedings of the 2nd Workshop on Noisy User-generated Text
(WNUT), pp. 102-111. 2016.
[3] da Costa Albuquerque, Fábio, Marco A. Casanova, Jose Antonio F. de
Macedo, Marcelo Tilio M. de Carvalho, and Chiara Renso. “A proactive
application to monitor truck leets,” In 2013 IEEE 14th International
Conference on Mobile Data Management, vol. 1, pp. 301-304. IEEE, 2013.
[4] D. McDonald, H. Chen, and R. Schumaker. “Transforming Open-Source
Documents to Terror Networks: The Arizona TerrorNet,” InAAAI Spring
Symposium: AI Technologies for Homeland Security, pp. 62-69, 2005.
[5] C.P. Wei, and Y.H. Lee. “Event detection from online news documents
for supporting environmental scanning,” Decision Support Systems 36,
pp. 385-401, 2004.
[6] M.H. Steinberg. “Clinical trials in sickle cell disease: adopting
the combination chemotherapy paradigm,” American Journal of
Hematology83, no. 1, pp. 1-3, 2008.
[7] S. Xiong, K. Wang, D. Ji, B. Wang. “A short text sentiment-topic model for
product reviews,” Neurocomputing 297, pp. 94-102, 2018.
[8] Abbasi, Ahmed, Stephen France, Zhu Zhang, and Hsinchun Chen.
“Selecting attributes for sentiment classiication using feature relation
networks,”IEEE Transactions on Knowledge and Data Engineering23, no.
3, pp. 447-462, 2010.
[9] Aggarwal, Charu C. “Machine Learning for Text: An Introduction,”
InMachine Learning for Text, pp. 1-16. Springer, Cham, 2018.
[10] Ahmed, Sajid, Asif Mahbub, Farshid Rayhan, Rafsan Jani, Swakkhar
Shatabda, and Dewan Md Farid. “Hybrid methods for class imbalance
learning employing bagging with sampling techniques,” In 2017 2nd
International Conference on Computational Systems and Information
Technology for Sustainable Solution (CSITSS), pp. 1-5. IEEE, 2017.
[11] Alcalá-Fdez, Jesús, Luciano Sanchez, Salvador Garcia, Maria Jose del
Jesus, Sebastian Ventura, Josep Maria Garrell, José Otero et al. “KEEL:
a software tool to assess evolutionary algorithms for data mining
problems,”Soft Computing13, no. 3, pp. 307-318, 2009.
[12] Armanfard, Narges, James P. Reilly, and Majid Komeili. “Local feature
selection for data classiication,” IEEE Transactions on Pattern Analysis
and Machine Intelligence38, no. 6, pp. 1217-1227, 2015.
[13] Bahassine, Said, Abdellah Madani, and Mohamed Kissi. “An improved
Chi-sqaure feature selection for Arabic text classiication using decision
tree,” In2016 11th International Conference on Intelligent Systems: Theories
and Applications (SITA), pp. 1-5. IEEE, 2016.
[14] Cao, Peng, Dazhe Zhao, and Osmar Zaiane. “An optimized cost-
sensitive SVM for imbalanced data learning,” InPaciic-Asia conference
on knowledge discovery and data mining, pp. 280-292. Springer, Berlin,
Heidelberg, 2013.
[15] Chen, Jingnian, Houkuan Huang, Shengfeng Tian, and Youli u. “Feature
selection for text classiication with Naïve Bayes,”Expert Systems with
Applications36, no. 3, pp. 5432-5435, 2009.
[16] S. Kumar, Ravishankar, and S. Verma. “Context Aware Dynamic
Permission Model: A Retrospect of Privacy and Security in Android
System,” in 2018 International Conference on Intelligent Circuits and
Systems , IEEE Xplore, Phagwara, India, pp. 324-329, 2018.
[17] T. Sabbah, A. Selamat, M.H. Selamat, F.S. Al-Anzi, E.H. Viedma, O. Krejcar,
and H. Fujita. “Modiied frequency-based term weighting schemes for
text classiication,” Applied Soft Computing 58, pp. 193–206, 2017.
[18] B. Vijayalakshmi, K. Ramar, NZ Jhanjhi, S. Verma, M. Kaliappan, et.al. “An
Attention Based Deep Learning Model For Trafic Flow Prediction Using
Spatio Temporal Features Towards Sustainable Smart City,” International
Journal of Communication Systems, 34, pp. 1-14 ,2020.
[19] S. Schmidt, S. Schnitzer, and C. Rensing. “Text classiication based ilters
for a domain-speciic search engine,” Computers in Industry 78, pp. 70–
79, 2016.
[20] Y. Liu, H.T. Loh, and A. Sun. “Imbalanced text classiication: A term
weighting approach,” Expert System Applications 36, pp. 690–701, 2013.
[21] Ghosh, Samujjwal, and Maunendra Sankar Desarkar. “Class speciic
TF-IDF boosting for short-text classiication: Application to short-texts
generated during disasters,” In Companion Proceedings of the The Web
Conference 2018, pp. 1629-1637. 2018.
[22] Dal Pozzolo, Andrea, Giacomo Boracchi, Olivier Caelen, Cesare Alippi,
and Gianluca Bontempi. “Credit card fraud detection: a realistic modeling
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 50 -
and a novel learning strategy,”IEEE Transactions on Neural Networks and
Learning Systems29, no. 8, pp. 3784-3797, 2017.
[23] Das, Sanjiv Ranjan. “Text and context: Language analytics in
inance,”Foundations and Trends® in Finance8, no. 3, pp. 145-261, 2014.
[24] I. Batra, S. Verma and Kavita, and M. Alazab. “A Lightweight IoT based
Security Framework for Inventory Automation Using Wireless Sensor
Network,” International Journal of Communication Systems 33, pp.1-
16, 2019.
[25] Elagamy, Mazen Nabil, Clare Stanier, and Bernadette Sharp. “Stock
market random forest-text mining system mining critical indicators
of stock market movements,” In 2018 2nd International Conference on
Natural Language and Speech Processing (ICNLSP), pp. 1-8. IEEE, 2018.
[26] García, Salvador, and Francisco Herrera. “Evolutionary undersampling
for classiication with imbalanced datasets: Proposals and
taxonomy,”Evolutionary Computation17, no. 3, pp. 275-306, 2009.
[27] Ghanem, Amal S., Svetha Venkatesh, and Geoff West. “Multi-class pattern
classiication in imbalanced data,” In2010 20th International Conference
on Pattern Recognition, pp. 2881-2884. IEEE, 2010.
[28] Gomez, Juan Carlos, and Marie-Francine Moens. “PCA document
reconstruction for email classiication,”Computational Statistics & Data
Analysis56, no. 3, pp. 741-751, 2012.
[29] Granitto, Pablo M., Cesare Furlanello, Franco Biasioli, and Flavia Gasperi.
“Recursive feature elimination with random forest for PTR-MS analysis
of agroindustrial products,” Chemometrics and Intelligent Laboratory
Systems83, no. 2, pp. 83-90, 2006.
[30] He, Haibo, and Edwardo A. Garcia. “Learning from imbalanced
data,”IEEE Transactions on Knowledge and Data Engineering21, no. 9, pp.
1263-1284, 2009.
[31] Jeatrakul, Piyasak, and Kok Wai Wong. “Enhancing classiication
performance of multi-class imbalanced data using the OAA-DB
algorithm,” InThe 2012 International Joint Conference on Neural Networks
(IJCNN), pp. 1-8. IEEE, 2012.
[32] Jin, Xin, Anbang Xu, Rongfang Bie, and Ping Guo. “Machine learning
techniques and chi-square feature selection for cancer classiication
using SAGE gene expression proiles,” In International Workshop on
Data Mining for Biomedical Applications, pp. 106-115. Springer, Berlin,
Heidelberg, 2006.
[33] H. Kaur, H.S. Pannu, and A.K. Malhi. “A systematic review on imbalanced
data challenges in machine learning: Applications and solutions,”ACM
Computing Surveys (CSUR)52, no. 4, pp. 1-36, 2019.
[34] L. Khreisat. “Arabic Text Classiication Using N-Gram Frequency Statistics
A Comparative Study,”In Conference on Data Mining (DMIN2006), pp.
78-82, 2006.
[35] S.B. Kotsiantis. “Decision trees: a recent overview,”Artiicial Intelligence
Review39, no. 4, pp. 261-283, 2013.
[36] B. Krawczyk. “Learning from imbalanced data: open challenges and future
directions,”Progress in Artiicial Intelligence5, no. 4, pp. 221-232, 2016.
[37] I. Batra, S. Verma, Kavita, U. Ghosh, J. J. P. C. Rodrigues, et al. “Hybrid
Logical Security Framework for Privacy Preservation in the Green
Internet of Things,” MDPI-Sustainability 12, no. 14, pp. 5542, 2020.
[38] J. Lee, I. Yu, J. Park, D.W. Kim. “Memetic feature selection for multilabel
text categorization using label frequency difference,” Information
Sciences485, pp. 263-280, 2019.
[39] G. Lemaître, F. Nogueira, and C.K. Aridas. “Imbalanced-learn: A
python toolbox to tackle the curse of imbalanced datasets in machine
learning,”The Journal of Machine Learning Research18, no. 1, pp. 559-
563, 2017.
[40] Jing, Li-Ping, Hou-Kuan Huang, and Hong-Bo Shi. “Improved feature
selection approach TFIDF in text mining,” InProceedings. International
Conference on Machine Learning and Cybernetics, vol. 2, pp. 944-946.
IEEE, 2002.
[41] G. Liang, C. Zhang. “A comparative study of sampling methods and
algorithms for imbalanced time series classiication,” In Australasian
Joint Conference on Artiicial Intelligence, pp. 637-648. Springer, Berlin,
Heidelberg, 2012.
[42] M. A. Jan, B. Dong, S. R. U. Jan, Z. Tazzn, S. Verma, et al. “A Comprehensive
Survey on Machine Learning-based Big Data Analytics for IoT-enabled
Smart Healthcare System,” Mobile Networks and Applications 26, pp.234-
252, Springer, 2021.
[43] P. Liu, X. Qiu, and H. Xuanjing. “Recurrent neural network for text
classiication with multi-task learning,” In Proceedings of the Twenty-
Fifth International Joint Conference on Artiicial Intelligence, pp.2873-
2879, 2016.
[44] X. Liu, Q. Li, and Z. Zhou. “Learning imbalanced multi-class data with
optimal dichotomy weights,” In2013 IEEE 13th International Conference
on Data Mining, pp. 478-487. IEEE, 2013.
[45] R.J. Lyon, J.M. Brooke, J.D. Knowles, and B.W. Stappers. “Hellinger
distance trees for imbalanced streams,” in 2014 22nd International
Conference on Pattern Recognition, pp. 1969-1974. IEEE, 2014.
[46] D. Fatta, Giuseppe, A. Fiannaca, R. Rizzo, A. Urso, M. R. Berthold, and S.
Gaglio. “Context-Aware Visual Exploration of Molecular Datab,” InSixth
IEEE International Conference on Data Mining-Workshops (ICDMW’06),
pp. 136-141. IEEE, 2006.
[47] A. Makazhanov, and D. Raiei, “Predicting the political preference of
Twitter users,” Social Network Analysis and Mining - ASONAM ’13, pp.
298–305, 2013.
[48] K. Mathew, and B. Issac. “Intelligent spam classiication for mobile text
message,” In Proceedings of 2011 International Conference on Computer
Science and Network Technology, vol. 1, pp. 101-105. IEEE, 2011 .
[49] C. Zhang, J. Bi, S. Xu, E. Ramentol, G. Fan, B. Qiao, and H. Fujita. “Multi-
Imbalance: An open-source software for multi-class imbalance learning,”
Knowledge Based System 174, pp. 137–143, 2019.
[50] A. Mazyad, F. Teytaud, and C. Fonlupt. “A comparative study on term
weighting schemes for text classiication”, in Lecture Notes in Computer
Science, Springer Verlag, pp. 100–108, 2018.
[51] A. Moreo, A. Esuli, and F. Sebastiani. “Distributional random
oversampling for imbalanced text classiication,” In Proceedings of the
39th International ACM SIGIR conference on Research and Development in
Information Retrieval, pp. 805-808, 2016.
[52] A. Onan, S. Korukoǧlu, and H. Bulut. “Ensemble of keyword extraction
methods and classiiers in text classiication,” Expert System Applications
57, pp. 232–247, 2016.
[53] N.C. Oza, and S. J. Russell. “Online bagging and boosting,” InInternational
Workshop on Artiicial Intelligence and Statistics, pp. 229-236., 2001.
[54] A. Özçift. “Random forests ensemble classiier trained with data
resampling strategy to improve cardiac arrhythmia diagnosis,”Computers
in Biology and Medicine41, no. 5, pp. 265-271, 2011.
[55] V.N. Phu, V.T.N. Tran, V.T.N. Chau, N.D. Dat, and K.L.D. Duy. “A decision
tree using ID3 algorithm for English semantic analysis,” International
Journal of Speech Technology20, no. 3, pp. 593-613, 2017.
[56] T. Pranckevičius, and V. Marcinkevičius. “Comparison of naive bayes,
random forest, decision tree, support vector machines, and logistic
regression classiiers for text reviews classiication,” Baltic Journal of
Modern Computing5, no. 2, pp. 221, 2017.
[57] M. Raza, F.K. Hussain, O.K. Hussain, M. Zhao, and Z. ur Rehman. “A
comparative analysis of machine learning models for quality pillar
assessment of SaaS services by multi-class text classiication of users’
reviews,” Future Generation Computer Systems 101, pp. 341–371, 2017.
[58] F. Khan, A. Shahnazir, N. Ayazsb, S. Khan, S. Verma, and Kavita.
“A Resource Eficient hybrid Proxy Mobile IPv6 extension for Next
Generation IoT Networks,” IEEE Internet of Things Journal, 2021, 10.1109/
JIOT.2021.3058982
[59] A. P. Singh, A. K. Luhach, S. Agnihotri, N. R. Sahu, D. S. Roy, NZ Jhanjhi,
S. Verma, Kavita, and U. Ghosh. “A Novel Patient-Centric Architectural
Framework for Blockchain-Enabled Healthcare Applications,” IEEE-
Transaction on Industrail Informatics 17, no. 8, pp. 5779 – 5789, 2020,
10.1109/TII.2020.3037889.
[60] R.E. Schapire, Y. Singer, and A. Singhal. “Boosting and Rocchio applied to
text iltering,” InProceedings of the 21st annual international ACM SIGIR
conference on Research and development in information retrieval, pp. 215-
223. 1998.
[61] R.P. Schumaker, and H. Chen. “Textual analysis of stock market
prediction using breaking inancial news: The AZFin text system,”ACM
Transactions on Information Systems (TOIS)27, no. 2, pp. 1-19, 2009.
[62] R.A. Stein, P.A. Jaques, and J.F. Valiati. “An analysis of hierarchical text
classiication using word embeddings,” Information Sciences 471, pp.
216–232, 2019.
[63] S. Tan. “Neighbor-weighted K-nearest neighbor for unbalanced text
corpus,” Expert System Applications 28, pp. 667–671, 2005.
[64] H. Tayyar Madabushi, E. Kochkina, and M. Castelle. “Cost-Sensitive
Special Issue on Artificial Intelligence in Economics, Finance and Business
- 51 -
BERT for Generalisable Sentence Classiication on Imbalanced Data,”
arXiv preprint arXiv:2003.11563, pp. 125–134, 2020.
[65] C.F. Tsai, W.C. Lin, Y.H. Hu, and G.T. Yao. “Under-sampling class
imbalanced datasets by combining clustering analysis and instance
selection,” Information Sciences 477, pp. 47–54, 2019.
[66] A.K. Uysal, and S. Gunal. “A novel probabilistic feature selection method
for text classiication,” Knowledge-Based Systems 36, pp. 226–235, 2012.
[67] B. Verma, and A. Rahman. “Cluster-oriented ensemble classiier: Impact
of multicluster characterization on ensemble classiier learning,” IEEE
Transactions on Knowledge and Data Engineering24, no. 4, pp. 605–618, 2012.
[68] M.K. Verma, D.K. Xaxa, and S. Verma. “DBCS: density based cluster
sampling for solving imbalanced classiication problem,” In 2017
International conference of Electronics, Communication and Aerospace
Technology (ICECA), vol. 1, pp. 156-161. IEEE, 2017.
[69] G. Yang, M. A. Jan, A. U. Rehman, M. Babar, and M. M. Aimal.
“Interoperability and Data Storage in Internet of Multimedia Things:
Investigating Current Trends, Research Challenges and Future
Directions,” IEEE Access 8, pp. 124382 – 124401, 2020.
[70] V. Dogra. “Banking news-events representation and classiication with
a novel hybrid model using DistilBERT and rule-based features,”Turkish
Journal of Computer and Mathematics Education (TURCOMAT)12, no. 10,
pp. 3039-3054, 2021.
[71] J. Yan, B. Zhang, N. Liu, S. Yan, Q. Cheng, W. Fan, Q. Yang, W. Xi, and Z.
Chen. “Effective and eficient dimensionality reduction for large-scale
and streaming data preprocessing,” IEEE Transactions on Knowledge and
Data Engineering18, no. 3, pp. 320–332, 2006.
[72] J. Yang, Y. Liu, X. Zhu, Z. Liu, and X. Zhang. “A new feature selection
based on comprehensive measurement both in inter-category and intra-
category for text categorization,” Information Processing Management 48,
pp. 741–754, 2012.
[73] K. Yang, Z. Yu, S. Member, X. Wen, W. Cao, S. Member, C.L.P. Chen, H.
Wong, and J. You. “Hybrid Classiier Ensemble for Imbalanced Data,”
IEEE Transactions on Neural Networks and Learning Systems31, no. 4, pp.
1–14, 2019.
[74] H. Zhang, and M. Li. “RWO-Sampling: A random walk over-sampling
approach to imbalanced data classiication,” Information Fusion 20, pp.
99–116, 2014.
[75] A. S. Ashour, S. Beagum, N. Dey, A. S. Ashour, D. S. Pistolla, , G. N.
Nguyen,... and F. Shi. “Light microscopy image de-noising using
optimized LPA-ICI ilter,” Neural Computing and Applications 29, no. 12,
pp. 1517-1533, 2018.
[76] S. Doss, J. Paranthaman, S. Gopalakrishnan, A. Duraisamy, S. Pal et
al. “Memetic optimization with cryptographic encryption for secure
medical data transmission in iot-based distributed systems,”Computers,
Materials & Continua 66, no.2, pp. 1577–1594, 2021.
[77] D. N. Le. “A new ant algorithm for optimal service selection with end-
to-end QoS constraints,”Journal of Internet Technology 18, no.5, pp. 1017-
1030, 2017.
[78] Z. Zhang, B. Krawczyk, S. Garcìa, A. Rosales-Pérez, and F. Herrera,
“Empowering one-vs-one decomposition with ensemble learning for multi-
class imbalanced data,” Knowledge Based System 106, pp. 251–263, 2016.
[79] Z. Sabir, K. Nisar, M. A. Z. Raja, M. R. Haque, M. Umar, A. A. A. Ibrahim,
and D. N. Le. “IoT technology enabled heuristic model with Morlet
wavelet neural network for numerical treatment of heterogeneous
mosquito release ecosystem,”IEEE Access9, pp. 132897-132913, 2021.
[80] T. Zhu, Y. Lin, and Y. Liu. “Synthetic minority oversampling technique for
multiclass imbalance problems,” Pattern Recognition 72, pp. 327–340, 2017.
Varun Dogra
Varun Dogra has been pursuing a Ph.D. in Computer
Applications at Lovely Professional University, Phagwara,
Punjab, India. He has Bachelor in Science and Masters
in Computer Applications. He has also been working as
Assistant Professor in the School of Computer Science
and Engineering, Lovely Professional University. He has
to have 14 years of experience in teaching/ industry. He
has published papers in reputed journals and presented papers in International
conferences. He has also been reviewed research papers of Scopus/ WoS
indexed journals. His area of research covers Articial Intelligence, Natural
Language Processing, Data Science, and Financial Markets.
Sahil Verma
Sahil Verma (Senior Member IEEE, ACM, IAENG) is Ph.
D in Computer Science and Engineering. He is an Associate
Professor and (A.) Directior in Chandigarh University,
Mohali, India. He has published many research articles in
reputed journals/publishers like IEEE, Wiley, Springer,
ACM, Elsevier, MDPI etc. He has published papers
in reputed top-cited journals like IEEE Transaction in
Industrial Informatics, IEEE Transaction on Network Science and Engineering,
IEEE Internet of Things Journals, ACM Transaction on Internet Technology,
CMC, IEEE Access, MONET Elsevier, HCIS Springer, MTAP Springer, MDPI
Sensors, Symmetry and many more. He is reviewer of top-cited journals like IEEE
Transaction on Intelligent Transport Systems, IEEE Transactions on Network
Science and Engineering, IEEE Access, Neural Computing and Applications
Springer, Human-centric Computing and Information Sciences Springer,
Mobile Networks and Applications Springer, Journal of Information Security
and Applications Elsevier, Mobile Information Systems Hindawi, International
Journal of Communication Systems Wiley, Security and Communication
Networks Hindawi etc. Dr. Verma is also had professional membership of many
reputed organisations like IEEE, ACM, IAENG. His tenure led to an overall
Excellence in Education, Research, Infrastructure and Systemic Development of
Organization. His current focus is to enhance the Quality of Education through
Strategic Quality Initiatives. He has visited many countries like: Austria, Czech
Republic, Germany, Switzerland, France, Italy and Thailand for exploring
research and development, establishment of labs and for the collaboration with
foreign universities (students exchange programs, faculty exchange programs etc.
Kavita Verma
Kavita Verma is Ph. D in Computer Science and Engineering.
She is an Associate Professor at Chandigarh University,
Mohali, India. She has published papers in reputed journals like
IEEE Transaction in Industrial Informatics, IEEE Transaction
on Network Science and Engineering, IEEE Internet of Things
Journals, ACM Transaction on Internet Technology, CMC,
IEEE Access, MONET Elsevier, HCIS Springer, MTAP
Springer, MDPI Sensors, Symmetry and many more. She is also a reviewer of top-cited
journals like IEEE Transaction on Intelligent Transport Systems, IEEE Transactions
on Network Science and Engineering, IEEE Access, Neural Computing, and
Applications Springer, Human-centric Computing and Information Sciences Springer,
Mobile Networks and Applications Springer, Journal of Information Security and
Applications Elsevier, Mobile Information Systems Hindawi, International Journal
of Communication Systems Wiley, Security and Communication Networks Hindawi,
etc. Dr. Kavita Verma has professional membership of many reputed organizations
like SMIEEE, MACM, MIAENG, MISCA.
Noor Zaman Jhanjhi
Noor Zaman Jhanjhi (NZ Jhanjhi) is currently working as
Associate Professor, Director Center for Smart society 5.0
[CSS5], and Cluster Head for Cybersecurity cluster, at School
of Computer Science and Engineering, Faculty of Innovation
and Technology, Taylor’s University, Malaysia. He is
supervising a great number of Postgraduate students, mainly
in cybersecurity for Data Science. The cybersecurity research
cluster has extensive research collaboration globally with several institutions
and professionals. Dr Jhanjhi is Associate Editor and Editorial Assistant Board
for several reputable journals, including IEEE Access Journal, PeerJ Computer
Science, PC member for several IEEE conferences worldwide, and guest editor
for the reputed indexed journals. Active reviewer for a series of top tier journals
has been awarded globally as a top 1% reviewer by Publons (Web of Science).
He has been awarded as outstanding Associate Editor by IEEE Access for the year
2020. He has high indexed publications in WoS/ISI/SCI/Scopus, and his collective
research Impact factor is more than 350 points as of the rst half of 2021. He has
international Patents on his account, edited/authored more than 30 plus research
books published by world-class publishers. He has great experience supervising
and co-supervising postgraduate students. An ample number of PhD and Master
students graduated under his supervision. He is an external PhD/Master thesis
examiner/evaluator for several universities globally. He has completed more than 22
international funded research grants successfully. He has served as Keynote speaker
for several international conferences, presented several Webinars worldwide,
chaired international conference sessions. His research areas include Cybersecurity,
IoT security, Wireless security, Data Science, Software Engineering, UAVs.
International Journal of Interactive Multimedia and Artificial Intelligence, Vol. 7, Nº3
- 52 -
Uttam Ghosh
Uttam Ghosh is currently working as Associate Professor
of Cybersecurity in Meharry School of Applied Computer
Science, Nashville, TN, USA. He has been over 10 years of
research and development experience in secure wireless and
wired communications, Software dened networking, CPS
Security. His area of research covers multiple domains like
Cyber Physical system Security, Mobile Ad hoc Networks,
Wireless Sensor Networks, Software-Dened Networking, Cloud Computing,
Distributed Algorithms, and Internet of Things(IoT). He has published many
research articles in reputed journals/publishers. He is also a reviewer of top-
cited journals. He has a professional membership of reputed organizations like
SMIEEE, Sigma Xi, ACM, IEEE, AAAS, ASEE.
Dac-Nhuong Le
Dac-Nhuong Le has an MSc and PhD in computer science
from Vietnam National University, Vietnam in 2009,
and 2015, respectively. He is an Associate Professor
on Computer Science, Deputy Head of the Faculty of
Information Technology, Haiphong University, Vietnam.
He has a total academic teaching experience of 20+ years
in computer science. He has more than 80+ publications in
the reputed international conferences, journals, and book chapter contributions
(Indexed by SCIE, SSCI, ESCI, Scopus). His areas of research are in the eld
of intelligence computing, multi-objective optimization, network security,
cloud computing, virtual reality/argument reality. Recently, he has been on
the technique program committee, the technique reviews, the track chair for
international conferences under Springer-ASIC/LNAI/CISC Series. Presently,
he is serving on the editorial board of international journals and edited/authored
20+ computer science books published by Springer, Wiley, CRC Press, Bentham
Publishers.
