Automatically Generated Classifiers for Opinion Mining with

Different Term Weighting Schemes

Shakhnaz Akhmedova

, Eugene Semenkin

and Roman Sergienko

Institute of Computer Science and Telecommunications, Siberian State Aerospace University, Krasnoyarsk, Russia

Institute of Communications Engineering, Ulm University, Ulm, Germany

Keywords: Opinion Mining, Support Vector Machine, Neural Networks, Bio-Inspired Algorithms.

Abstract: Automatically generated classifiers using different term weighting schemes for Opinion Mining are

presented. New collective nature-inspired self-tuning meta-heuristic for solving unconstrained and

constrained real- and binary-parameter optimization problems called Co-Operation of Biology Related

Algorithms was developed and used for classifiers design. Three Opinion Mining problems from DEFT’07

competition were solved by proposed classifiers. Also different weighting schemes were used for data

processing. Obtained results were compared between themselves and with results obtained by methods

which were proposed by other researchers. As the result workability and usefulness of designed classifiers

were established and best data processing approach for them was found.

1 INTRODUCTION

Opinion Mining (also called sentiment analysis) is

the process of determining the attitude of a speaker

or a writer with respect to some topic or the overall

context of a document (Pang and Lee, 2008). The

attitude may be some person’s judgment or

evaluation, affective or emotional state, or the

intended emotional communication.

Most opinion mining algorithms use simple

terms to express sentiment about a product or

service (for example, a “positive” or “negative”

review). However, cultural factors, linguistic

nuances and differing contexts make it extremely

difficult to turn a string of written text into a simple

pro or con sentiment.

Nowadays various techniques are developed for

solving opinion mining problems, for example latent

semantic analysis, "bag of words" and other machine

learning algorithms (Pang, Lee and Vaithyanathan,

2002). But it is well known that the way that

documents are represented influences on the

performance of the text classification (opinion

mining problems can also be considered as text

categorization problems) algorithms (Youngjoong

Ko, 2012).

In this work artificial neural networks and

support vector machines automatically generated by

collective bionic algorithm and its modifications are

decribed. They were used with eleven different text

preprocessing techniques for solving three opinion

mining problems which were taken from the

DEFFT’07 competition. The best combinations (text

preprocessing for problems) for classifiers were

found.

In the second section proposed algorithms are

described. Term weighting schemes are shown in the

third section. Next experimntal results are presented

and some conclusions are given.

2 AUTOMATICALLY

GENERATED CLASSIFIERS

2.1 Co-Operation of Biology Related

Algorithms

A new collective bionic method called Co-Operation

of Biology Related Algorithms (COBRA) was

introduced in (Akhmedova and Semenkin, 2013).

It’s based on the co-operative work of five well

known heuristics: Particle Swarm Optimization

(PSO) (Kennedy and Eberhart, 1995), Wolf Pack

Search (WPS) (Yang, Tu and Chen, 2007), Firefly

Algorithm (FFA) (Yang, 2009), Cuckoo Search

Algorithm (CSA) (Yang and Deb, 2009) and Bat

Algorithm (BA) (Yang, 2010).

845

Akhmedova S., Semenkin E. and Sergienko R..

Automatically Generated Classiﬁers for Opinion Mining with Different Term Weighting Schemes.

DOI: 10.5220/0005148508450850

In Proceedings of the 11th International Conference on Informatics in Control, Automation and Robotics (ASAAHMI-2014), pages 845-850

ISBN: 978-989-758-040-6

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

The proposed approach is a self-tuning algorithm:

the population size for each method can increase or

decrease during the program run. Besides,

populations communicate with each other: they

exchange individuals after a given number of

objective function evaluations.

The proposed heuristic’s workability and

usefulness were established by its testing on various

optimization problems (Liang et al., 2012). Also it

was established that COBRA outperforms its

component-algorithms and exhibits a competitive

level of performance compared to alternative

optimization techniques.

The COBRA heuristic was originally developed

for solving real-parameter unconstrained

optimization problems. Later it was modified for

solving not only real but also binary-parameter

constrained and unconstrained optimization

problems.

Using the technique described in (Kennedy and

Eberhart, 1997), the COBRA binary modification

(COBRA-b) was developed. COBRA was adapted to

search in binary spaces by applying a sigmoid

transformation which also was taken from (Kennedy

and Eberhart, 1997) to the velocity (PSO, BA) and

coordinates (FFA, CSA, WPS) to compress them

into a range [0, 1] and force the component values of

the individual position to be 0’s or 1’s.

Then COBRA’s modification for solving

constrained optimization problems was also

developed. Three constraints handling methods were

used for this purpose: dynamic penalties (Eiben and

Smith, 2003), Deb’s rule (Deb, 2000) and the

technique described in (Liang, Shang and Li, 2010).

Method proposed in (Liang, Shang and Li, 2010)

was implemented to PSO-component of COBRA; at

the same time other components were modified by

implementing Deb’s rule followed by calculating

function values using dynamic penalties. The new

algorithm was called COBRA-c (Akhmedova and

Semenkin, 2013).

The performance of both modifications was

evaluated with a set of various test functions. For

example, 18 scalable benchmark functions provided

by the CEC 2010 competition and a special session

on single objective constrained real-parameter

optimization (Mallipeddi, Suganthan, 2009) were

used for heuristic COBRA-c testing. The constrained

modification of COBRA was compared with

algorithms that took part in the CEC’2010

competition. It was established that COBRA-b and

COBRA-c work successfully and are sufficiently

reliable. Besides, the proposed approach for

constrained optimization problems is superior to 3-4

of the 14 winning methods from this competition.

And finally, COBRA’s modifications outperform all

of its component algorithms.

2.2 ANN-Based Classifiers

The feed-forward artificial neural network (ANN)

models have three primary components: the input

data layer, the hidden layer(s) and the output layer.

Each of these layers contains nodes and these nodes

are connected to nodes at adjacent layer(s). Also

each node has its own activation function. So the

number of hidden layers, the number of nodes

(neurons), and the type of activation function on

each node will be denoted as “ANN’s structure”.

Before solving a classification problem by neural

network its structure should be chosen.

Besides, nodes in network are interconnected and

each connection has a weight coefficient; the

number of these coefficients depends on the solving

problem (number of inputs) and the number of

hidden layers and nodes. Thus, networks with a

more or less complex structure usually have many

weight coefficients which should be adjusted.

The neural networks’ structure design and the

tuning of their weight coefficients are considered as

the solving two unconstrained optimization

problems: the first one with binary variables and the

second one with real-valued variables. The type of

variables depends on the representation of the

ANN’s structure and coefficients.

For this work we set a maximum number of

hidden layers and a maximum number of neurons on

each hidden layer (both equal to 5), so that the

maximum number of hidden neurons is equal to 25.

We could have chosen a larger number of layers and

nodes, but our aim was to show that an even network

with a relatively small structure can show good

results if it is tuned with effective optimization

techniques.

Each node was represented by a binary string of

length 4. If the string consisted only of zeros

(“0000”) then this node wouldn’t exist in ANN. So,

the whole structure of the neural network was

represented by a binary string of length 100 (25x4),

and each 20 variables represented one hidden layer.

The number of input layers depended on the

problem in hand.

Also we used 15 different activation functions

for all nodes, for example, sigmoidal function, linear

function, hyperbolic tangent function and so on. For

determining which activation function would be

used on a given node, the integer corresponding to

its binary string was calculated. Namely, if some

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

846

neuron had the binary string “0110”, then the integer

was calculated in the following way: 0×2

+ 1×2

1×2

+ 0×2

= 6. So for this neuron we used the

sixth activation function from the 15 mentioned

above.

Thus we used the optimization method COBRA-

b for finding the best ANN’s structure and the

optimization method COBRA for the weight

coefficients adjustment of every structure.

2.3 SVM-based Classifiers

Support vector machines are linear classification

mechanisms, which represent examples from a

training set as points in space mapped so that the

examples of the separate categories are divided by a

clear gap that is as wide as possible (Vapnik and

Chervonenkis, 1974). New examples (from a test

set) are then mapped into the same space and

predicted to belong to a category based on which

side of the gap they fall. So, SVM-based classifiers

linearly divide examples from different classes.

SVM is based on the maximization of the

distance between the discriminating hyper-plane and

the closest examples. This maximization reduces the

so-called structural risk, which is related to the

quality of the decision function. The most

discriminating hyper-plane can be computed by

solving constrained the real-parameter optimization

problem described in (Vapnik and Chervonenkis,

1974).

However, instances from different categories are

not always linearly separable, and in this case SVM

(as a linear classifier) does not provide satisfying

classification results. One way to resolve this

problem is to map the data onto a higher dimension

space and then to use a linear classifier in that space.

The general idea is to map the original feature space

to a higher-dimensional feature space where the

training set is linearly separable. SVM provides an

easy and efficient way of performing this mapping

to a higher dimensional space, which is referred to

as the kernel trick (Boser, Guyon and Vapnik, 1992).

In this study we use COBRA and COBRA-c to

automatically design appropriate SVM-based

classifiers for the problem in hand.

3 TERM WEIGHTING SCHEMES

Generally, text documents are not classified as

sequences of symbols. They are usually transformed

into vector representation in so-called feature space

because most machine learning algorithms are

designed for vector space models. The document

mapping into the feature space remains a complex

non-trivial task.

Text pre-processing techniques can be

considered as term weighting schemes: calculating

the weight of each word. The term weighting

methods can be roughly divided into two groups:

supervised and unsupervised methods. Almost all of

them use the frequency of the term occurring.

Two of the most popular text pre-processing

approaches among researchers are the TF-IDF

technique (Salton and Buckley, 1988) and the

ConfWeight method (Soucy and Mineau, 2005). In

this study we used the above schemes and also the

term weighting method proposed in (Gasanova et al.,

2013) called C-values. It has some similarities with

the ConfWeight method, but has improved

computational efficiency: no morphological or stop-

word filtering before text pre-processing was used. It

means that the text pre-processing can be performed

without a human expert or linguistic knowledge and

that the text pre-processing technique doesn’t

depend on the language. All term weighting schemes

(binary pre-processing, 8 TF-IDF techniques,

ConfWeights, C-values) which were used are briefly

described in (Gasanova et al., 2014).

Thus we tried to determine which of the term

weighting schemes mentioned is the most useful for

solving Opinion Mining problems.

4 EXPERIMENTAL RESULTS

The DEFT07 (“Défi Fouille de Texte”) Evaluation

Package (Plate-forme AFIA 2007) has been used for

the application of algorithms and the comparison of

results. For the testing of the proposed approach

three corpora were used: “Books”, “Video games”

(later just “Games”) and “Debates in Parliament”

(later just “Debates”). Descriptions of corpora are

given in Table 1.

These corpora are divided into train (60%) and

test (40%) sets by the organizers of the DEFT’07

competition and this partition has been retained in

our study to be able to directly compare the

performance achieved using the methods developed

in this work with the algorithms of participants. The

train and test set parameters of all corpora are shown

in Table 2.

In order to apply the classification algorithms, all

words which appear in the train set have been

extracted. Then words have been brought to the

same letter case: dots, commas and other

punctuation marks have been removed. It should be

AutomaticallyGeneratedClassifiersforOpinionMiningwithDifferentTermWeightingSchemes

847

mentioned that no other information related to the

language or domain (no stop or ignore word lists)

has been used in the pre-processing.

Table 1: Test corpora.

Corpus Description

Marking

scale

Books

3000 commentaries

about books, films

and shows

0:unfavorable,

1:neutral,

2:favourable

Games

4000 commentaries

about video games

0:unfavorable,

1:neutral,

2:favourable

Debates

28800 interventions

by Representatives in

the French Assembly

0:against the

proposed law,

1:for it

Table 2: Corpora sizes.

Corpus Data set sizes

Classes

(train

set)

Classes

(test set)

Books

Train size = 2074

Test size = 1386

Vocabulary = 52507

0:309,

1:615, 2:

1150

0:207,

1:411, 2:

768

Games

Train size = 2537

Test size = 1694

Vocabulary = 63144

0:497,

1:1166,

2:874

0:332,

1:583,

2:779

Debates

Train size = 17299

Test size = 11533

Vocabulary = 59615

0:10400,

1:6899

0:6572,

1:4961

The F-score value (Van Rijsbergen, 1979) was

used for evaluating the obtained results. The F-score

depends on the “precision” and “recall” of each

criterion. The classification “precision” for each

class is calculated as the number of correctly

classified instances for a given class divided by the

number of all instances which the algorithm has

assigned for this class. “Recall” is the number of

correctly classified instances for a given class

divided by the number of instances that should have

been in this class.

From the viewpoint of optimization, the design

of ANN-based classifiers for the mentioned corpora

requires solving optimisation problems having from

115 to 120 real-valued variables for ANN weight

coefficients and 100 binary variables for the ANN

structure. For example, here is the structure of the

neural network with the best obtained result using

the C-values pre-processing scheme for the problem

“Books” which has five hidden layers, three neurons

on the first layer, four neurons on the third layer and

five neurons on the other layers:

 The first layer is (0000 0000 0011 1100 1100),

i.e. neurons with the 3rd and 12th activation

functions;

 The second layer is (0001 0111 1100 0111

1111), i.e., neurons with the 1st, 7th, 12th, and

15th activation functions;

 The third layer is (1011 0111 1110 1111 0000),

i.e., neurons with the 11th, 7th, 14th, and 15th

activation functions

 The fourth layer is (0001 1001 0100 1101

1111), i.e., neurons with the 1st, 9th, 4th, 13th,

and 15th activation functions;

 The fifth layer is (0011 0110 1011 0101 1110),

i.e., neurons with the 3rd, 6th, 11th, 5th and

15th activation functions.

The results for all text categorization problems

obtained by generated SVM-based and ANN-based

classifiers with different term weighting schemes are

presented in Tables 3, 4 and 5.

Table 3: Results obtained for corpus “Books”.

Books SVM ANN

C-values 0.619 0.585137

ConfWeight 0.588023

0.613048

Binary_SUM 0.558442 0.566378

TF-IDF_1 0.580087 0.554113

TF-IDF_2 0.563492 0.533189

TF-IDF_3_0.1 0.577201 0.518038

TF-IDF_3_0.5 0.576479 0.460317

TF-IDF_3_0.9 0.55772 0.505051

TF-IDF_4_0.1 0.549784 0.553719

TF-IDF_4_0.5 0.559163 0.550767

TF-IDF_4_0.9 0.561328 0.541913

Table 4: Results obtained for corpus “Games”.

Games SVM ANN

C-values 0.695772 0.691919

ConfWeight 0.645218

0.726551

Binary_SUM 0.681818 0.65368

TF-IDF_1 0.668831 0.642136

TF-IDF_2 0.661157 0.649351

TF-IDF_3_0.1 0.686541 0.678932

TF-IDF_3_0.5 0.65987 0.643579

TF-IDF_3_0.9 0.603896 0.627706

TF-IDF_4_0.1 0.691263 0.701299

TF-IDF_4_0.5 0.645218 0.678932

TF-IDF_4_0.9 0.657096 0.551948

So the best classification quality for the problem

“Books” is provided with the C-values approach for

text pre-processing and Support Vector Machine

generated with COBRA as a classification method.

This result is better than the one obtained by the

DEFT’07 participants although no term filtering has

been used in the text pre-processing. The second

best result for the “Books” corpora was obtained by

Artificial Neural Network generated by COBRA but

with the ConfWeight term weighting scheme.

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

848

Table 5: Results obtained for corpus “Debates”.

Debates SVM ANN

C-values 0.699905 0.704587

ConfWeight 0.714211 0.709269

Binary_SUM 0.641984 0.6471

TF-IDF_1 0.638429 0.640336

TF-IDF_2 0.64129 0.640683

TF-IDF_3_0.1 0.661406 0.645712

TF-IDF_3_0.5 0.668659 0.611983

TF-IDF_3_0.9 0.669323 0.6436

TF-IDF_4_0.1 0.663314 0.643284

TF-IDF_4_0.5 0.665135 0.601231

TF-IDF_4_0.9 0.660827 0.566375

For the problem “Games” ANN-based classifiers

showed better results than Support Vector Machines.

And again the best result achieved by SVM-based

classifiers was provided with the C-values pre-

processing technique, while the best result achieved

by ANN-based classifiers was with the ConfWeight

approach.

There is no significant difference between the

results obtained by neural networks with the C-

values term weighting scheme and the ConfWeight

method for the problem “Debates”. However ANN-

based classifiers showed the best results with the

text pre-processing technique ConfWeight for this

corpus. Support Vector Machines were more

successful while solving the given problem. But that

time the best result was achieved by SVM-based

classifiers also with the ConfWeight approach.

It was established that generally the best term

relevance scheme for SVM-based classifiers is the

C-values technique and the ConfWeight method was

more useful for neural networks.

Table 6 contains the best results obtained with

SVM-based and ANN-based classifiers

automatically generated by developed bionic

algorithms. There are also results obtained for the

best submission of other researchers for each corpus.

The results for each corpus were ranked and then the

total rank was evaluated as a mean value. So the best

results were obtained by the method with the

smallest total rank, and vice versa, and the worst

results were obtained by the method with the largest

total rank value.

It should be noted that in Table 6 only the best

combinations of text pre-processing methods for our

algorithms are presented. Thus the proposed

classification methods outperformed almost all

alternative approaches. In (Gasanova et al., 2014)

results obtained by k-nearest neighbours algorithm

(k varied from 1 to 15) were presented. The best F-

score achieved by k-NN classifiers (with k equal to

15) for corpus “Books” was equal to 0.5593 and for

Table 6: Results obtained for all corpora, comparison of

the performance.

Team or

method

Books

(rank1)

Games

(rank2)

Debates

(rank3)

Rank

J.-M. Torres-

Moreno (LIA)

0.603 (3) 0.784 (1) 0.720 (1) 1

G. Denhiere

(EPHE)

0.599 (4) 0.699 (6) 0.681 (6) 4

S. Maurel

(CELI France)

0.519 (8) 0.706 (5) 0.697 (5) 5

M. Vernier

(GREYC)

0.577 (5) 0.761 (3) 0.673 (7) 6

E. Crestan

(Yahoo ! Inc.)

0.529 (7) 0.673 (8) 0.703 (4) 7

M. Plantie

(LGI2P)

0.472

(10)

0.783 (2) 0.671 (9) 8

A.-P. Trinh

(LIP6)

0.542 (6) 0.659 (9) 0.676 (8) 9

M. Genereux

(NLTG)

0.464

(11)

0.626

(10)

0.569

(12)

E. Charton

(LIA)

0.504 (9)

0.619

(11)

0.616

(10)

A. Acosta

(Lattice)

0.392

(12)

0.536

(12)

0.582

(11)

SVM+COBRA 0.619 (1) 0.696 (7) 0.714 (2) 3

ANN+COBRA 0.613 (2) 0.727 (4) 0.709 (3) 2

corpus “Debates” it was equal to 0.695. But for

problem “Games” k-NN algorithm outperformed

SVM-based classifiers generated by COBRA, its

result was 0.7203.

5 CONCLUSIONS

In this paper we have described a new meta-

heuristic, called Co-Operation of Biology Related

Algorithms, and introduced its modification for

solving unconstrained optimization problems with

binary variables (COBRA-b) and constrained

optimization problems with real-valued variables

(COBRA-c). The proposed algorithms’ usefulness

and workability were established by their testing on

sets of benchmark problems.

Then we used described optimization methods

for the automated design of ANN-based and SVM-

based classifiers. These approaches were applied to

three Opinion Mining problems which were taken

from the DEFT’07 competition. For that purpose

different text pre-processing techniques were used.

Solving these problems is equivalent to solving

big and hard optimization problems where objective

functions have many variables and are given in the

form of a computational program. The suggested

algorithms successfully solved all the problems of

designing classifiers with competitive performance.

AutomaticallyGeneratedClassifiersforOpinionMiningwithDifferentTermWeightingSchemes

849

A comparison with alternative classification

methods showed that SVM-based and ANN-based

classifiers designed by COBRA outperformed

almost all of them. This fact allows us to consider

the study results as confirmation of the reliability,

workability and usefulness of the algorithms in

solving real world optimization problems.

Having these appropriate tools for data mining,

we consider the following directions for the

approach development: the design of other types of

neural network models, the design of Support Vector

Machines with alternative kinds of kernel function,

the application to the design of fuzzy systems, the

improvement in optimization performance of

developed algorithms COBRA, COBRA-b and

COBRA-c.

ACKNOWLEDGEMENTS

Research is performed with the financial support of

the Ministry of Education and Science of the

Russian Federation within the federal R&D

programme (project RFMEFI57414X0037).

REFERENCES

Actes de l'atelier DEFT'07. Plate-forme AFIA 2007.

Grenoble, Juillet. http://deft07.limsi.fr/actes.php

Akhmedova, Sh., Semenkin, E., 2013. Co-Operation of

Biology related Algorithms. In IEEE Congress on

Evolutionary Computations. IEEE Publications.

Akhmedova, Sh., Semenkin, E., 2013. New optimization

metaheuristic based on co-operation of biology

related algorithms, Vestnik. Bulletine of Siberian

State Aerospace University. Vol. 4 (50).

Boser, B., Guyon, I., Vapnik, V., 1992. A training

algorithm for optimal margin classifiers. In The 5th

Annual ACM Workshop on COLT. ACM.

Deb, K., 2000. An efficient constraint handling method for

genetic algorithms, Computer methods in applied

mechanics and engineering. Vol. 186(2-4).

Eiben, A.E., Smith, J.E., 2003. Introduction to

evolutionary computation, Springer. Berlin.

Gasanova, T., Sergienko, R., Minker, W., Semenkin, E.,

Zhukov, E., 2013. A Semi-supervised Approach for

Natural Language Call Routing. In SIGDIAL 2013

Conference.

Gasanova, T., Sergienko, R., Akhmedova, Sh., Semenkin,

E., Minker, W., 2014. Opinion Mining and Topic

Categorization with Novel Term Weighting. In 5th

Workshop on Computational Approaches to

Subjectivity, Sentiment and Social Media Analysis,

Association for Computational Linguistics.

Kennedy, J., Eberhart, R., 1995. Particle Swarm

Optimization. In IEEE International Conference on

Neural Networks.

Kennedy, J., Eberhart, R., 1997. A discrete binary version

of the particle swarm algorithm. In World

Multiconference on Systemics, Cybernetics and

Informatics.

Liang, J.J., Qu, B.Y., Suganthan, P.N., Hernandez-Diaz,

A.G., 2012. Problem Definitions and Evaluation

Criteria for the CEC 2013 Special Session on Real-

Parameter Optimization. Technical Report,

Computational Intelligence Laboratory, Zhengzhou

University, Zhengzhou China, and Technical Report,

Nanyang Technological University, Singapore.

Liang, J.J., Shang Z., Li, Z., 2010. Coevolutionary

Comprehensive Learning Particle Swarm Optimizer.

In CEC’2010, Congress on Evolutionary

Computation. IEEE Publications.

Mallipeddi, R., Suganthan, P.N., 2009. Problem

Definitions and Evaluation Criteria for the CEC 2010

Competition on Constrained Real-Parameter

Optimization. Technical report, Nanyang

Technological University, Singapore.

Pang, B., Lee, L, 2008. Opinion Mining and Sentiment

Analysis, Now Publishers Inc. New-York.

Pang, B., Lee, L., Vaithyanathan, Sh., 2002. Thumbs up?

Sentiment Classification using Machine Learning

Techniques. In EMNLP, Conference on Empirical

Methods in Natural Language Processing.

Salton, G., Buckley, C., 1988. Term-weighting approaches

in automatic text retrieval, Information Processing and

Management. Vol. 24 (5).

Soucy, P., Mineau, G.W., 2005. Beyond TFIDF

Weighting for Text Categorization in the Vector Space

Model. In IJCAI’2005, The 19th International Joint

Conference on Artificial Intelligence.

Van Rijsbergen, C.J., 1979. Information Retrieval.

Butterworth, 2

edition.

Vapnik, V., Chervonenkis, A., 1974. Theory of Pattern

Recognition, Nauka. Moscow.

Yang, Ch., Tu, X., Chen, J., 2007. Algorithm of Marriage

in Honey Bees Optimization Based on the Wolf Pack

Search. In International Conference on Intelligent

Pervasive Computing.

Yang, X.S., 2009 Firefly algorithms for multimodal

optimization. In The 5th Symposium on Stochastic

Algorithms, Foundations and Applications.

Yang, X.S., 2010. A new metaheuristic bat-inspired

algorithm. Nature Inspired Cooperative Strategies for

Optimization, Studies in Computational Intelligence.

Vol. 284.

Yang, X.S., Deb, S., 2009. Cuckoo Search via Levy

flights. In World Congress on Nature & Biologically

Inspired Computing. IEEE Publications.

Youngjoong Ko, 2012. A study of term weighting

schemes using class information for text classification.

In SIGIR'12, The 35th Annual SIGIR Conference.

ACM.

ICINCO2014-11thInternationalConferenceonInformaticsinControl,AutomationandRobotics

850