Semantic Enrichment of Relevant Feature Selection Methods

for Data Mining in Oncology

Adriana Da Silva Jacinto

1,2

, Ricardo Da Silva Santos

2,3

and José Maria Parente De Oliveira

Department of Computer Science, Faculty of Technology, São José dos Campos, Brazil

Department of Computer Science, Aeronautics Institute of Technology, São José dos Campos, Brazil

Integrated Center of Biochemistry Research, University of Mogi das Cruzes, Mogi das Cruzes, Brazil

1 STAGE OF THE RESEARCH

Medical field is one of the most important for

society, in which several researches are performed.

One of the main concerns of responsible

organizations is cancer, which is a disease in which

abnormal cells divide uncontrollably and invade

other tissues, forming a malignant tumor in the most

of the cases (Jemal, 2011; NCI, 2014).

In 2025, the number of new cases of the disease

in the world is estimated around 19.3 million (Ferlay

et al., 2012; Bray et al., 2013), i.e., the number of

cases of the disease increases and the discovery of

the cure seems remote.

Due to that situation, several medical databases

of patients related to cancer are provided for

research and analysis (NCI, 2014). The goal is to

search improvements on the prognosis, diagnosis,

prevention, treatment for the disease, the actions and

decision-making by public health managers.

Use of data mining is presented as an aid to the

analysis of those data, because it is the exploitation

and analysis of databases through a variety of

statistical techniques and machine learning

algorithms for the discovery of rules and relevant

patterns (Tan et al., 2005).

The success of data mining is subordinate to the

selection of the most relevant features (Lin et al.,

2006). However, the correct selection of features

depends on methods and is highly related to the

knowledge of domain experts, because the meanings

of some features are not easily understood by a

person who is not familiar with the application

domain, complicating the insight about which of

them are useful for data mining (Kuo et al., 2007).

Generally, oncology specialists are not available

to help a data miner full time. Moreover, the medical

field has several terms and peculiarities, which may

make hard the selection of the semantically most

relevant features. Additionally, the analysis of the

importance of the meaning of each feature in a data

set with thousands of genes as features, for example,

would require too long time of a human expert,

making it nonviable.

The high number of features in databases is

pointed out by several researchers (Hall, 2000;

Freitas, 2001; Inbarani et al., 2007; Tan et al., 2009;

Mansingh et al., 2011) as one of the crucial

problems in data mining. Over time, in order to

remedy this problem, various selection methods

were developed to improve the functionality of the

pre-processing of the data (Ammu e Preeja, 2013;

Chahkandi, 2013).

As seen in this work, those methods only employ

statistical techniques to select the features and do not

capture the semantics of them. As result, the more

semantically relevant features are excluded for not

having statistical significance, while irrelevant

features are selected due to their extensive statistical

contribution.

Thus, this project presents a proposal of

capturing of the semantic importance of each feature

by computational manner. The proposal enriches

the traditional methods of feature selection by using

of Natural Language Processing, the NCI ontology,

WordNet and medical documents.

A prototype of this approach was implemented

and tested with five data sets related to cancer

patients. The results show that the use of semantic

improves the pre – processing by selecting of the

most relevant semantic features.

2 OUTLINE OF OBJECTIVES

This work aims to:

 Define an architecture for semantic enrichment

of feature selection filter methods, that

considers the semantics of the input features in

relation to predictable class;

 Implement the architecture and apply it in

Oncology;

Da Silva Jacinto A., Da Silva Santos R. and Parente De Oliveira J..

Semantic Enrichment of Relevant Feature Selection Methods for Data Mining in Oncology.

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

 Evaluate whether the proposed approach

presents improvements to the process of data

mining.

3 RESEARCH PROBLEM

According to Lee (2005), the problem of feature

selection can be examined from different

viewpoints, and one of which is the search for the

most relevant features, i.e., the searching for those

feature that bring more meaning to the obtained data

mining model. However, generally, the feature

selection methods use statistical techniques and do

not consider the semantics of each attribute.

Due to the fact of the feature selection methods

do not consider the semantic of features the

following problems can arise: a) elimination of

semantically relevant features that do not show their

importance in statistical data analysis; b)

consideration of irrelevant or redundant features due

to their high level of statistical contribution.

This work consider irrelevant feature regarding

knowledge to be extracted is the one, if removed,

causes no impact on data mining model, and

redundant feature is one whose value is already

represented by another.

Wu et al. (2011) highlight that data mining is a

complicated activity for beginners in the area,

resulting in the specification of several parameter

settings or selection of features that are syntactically

accepted, but are not semantically reasonable.

Also true there is no mechanism that can help

users on the semantics of configurations, it means

that the user has difficulty when trying to identify

which features must be selected for a given data

mining task, even after having a suggested list of

features given by some selection method (Wu et al.,

2011). That occurs, as made by Microsoft Visual

Studio (Microsoft, 2014), because methods of

selection of features do not consider the semantics

between features and provide a list of selected

features by statistical means, whose refinement is

left to the user, can leave him confused.

Presented the context, the research problem of

this paper is enunciated as follow:

Current feature selection methods do not

consider the semantics of each feature, and the

result is the selection of irrelevant or redundant

data, which increase the computational cost of the

operation and the generation of data mining models

with low quality.

4 STATE OF THE ART

Since 1990, a variety of feature selection methods

have been implemented using different approaches

and techniques. Nevertheless, this study sought to

identify the core of the approach used on several

feature selection methods in order to group them by

similarities.

Considered methods are only those of the filter

type, it means that their implementation and

execution are independent of the mining algorithm.

Thus, the main approaches for feature selection in

the literature are based on Entropy, Consistency,

Matrix Resources, Rough Sets, Similarity and

Correlation.

Entropy and others (the probability calculus,

symmetric uncertainty, information gain, mutual

information and frequency values) quantify the

disorder among the elements of a data set, with basis

on the calculation of probability (Cover and

Thomas, 1991). When a data set is homogeneous,

lower is the entropy among its elements.

Feature selection methods consider that high

entropy of an input attribute relative to a predictable

class shows its relevance. Main idea is that

variations in the values of a feature generate a

greater degree of uncertainty concerning what

happens in the prediction feature (Hall, 2000).

Examples of feature selection methods that

employ entropy and similar are: Focus and Focus-2

(Almuallim and Dietterich, 1991, 1992);

Correlation-based Feature Selection (CFS) (Hall,

2000); Based Fast Correlation Filter (FCBF) (Yu

and Liu, 2003); Interact (Zhao and Liu, 2007);

Information Theoretic-based Interact (IT-IN) (Deisy

et al, 2010.); TRS + Focus2 (Teruya, 2008);

Information Gain Attribute Ranking (Cover and

Thomas, 1991). Other approach adopted by the

feature selection methods is the calculation of

consistency, which refers to coherency. Whether

there are two tuples with the same values of input

feature, they must have equal values for the

prediction feature. If it does not, set up an example

of inconsistency (Dash and Liu, 2003; Liu and

Setiono, 1996). Thus, the subset of input features

which displays the lowest level of inconsistency will

be chosen by the selection method.

Following feature selection methods use the

calculation of consistency to elect the most relevant

input features: Las Vegas Filter (LVF) and

Consistency Subset Evaluation (CSE) (Liu and

Setiono, 1998); Information Theoretic-based Interact

(IT-IN) (Deisy et al, 2010.); Focus and Focus-2

(Almuallim and Dietterich, 1991, 1992) and Interact

SemanticEnrichmentofRelevantFeatureSelectionMethodsforDataMininginOncology

(Zhao and Liu, 2007).

Proceeding the description of the approaches

found, there is the use of matrix resource (matrix

SVD Laplacian matrix), which consists of the

decomposition of a data matrix into singular values,

calculating the cosine between two columns or

covariance matrix. Each column of the matrix is

considered as a vector. The aim is to modify the

original data space, identifying input features that

can be disregarded, due to its low contribution to the

set of features, measured by calculating the

eigenvalues (Pearson, 1901).

Use of matrix resources is checked out on the

following features selection methods: Spectral

Feature Selection (Spec) (Zhao et al, 2007.);

Laplacian Score (He et al, 2005.); Principal

Components Analysis (PCA) (Pearson, 1901).

Whereas a data set is formed by different

vectors, these vectors can be compared, establishing

between them a measure, which is calculated with

basis on a metric. The metric can be the cosine of

the resulting angle, the Euclidean distance or

Manhattan distance between the vectors and some

variations. This approach is called similarity or

correlation, and vectors can be tuples or features

(Meira Jr. and Zaki, 2009).

Use of similarity and correlation happens in the

following selection methods: Minimum-

Redundancy-Maximum-Relevance (mrMr) (Peng et

al, 2005.); Redundancy Demoting (RD) (Osl et al,

2009.); FCBF (Yu and Liu, 2003); ReliefF ReliefF-1

and-2 (Rendell and Kira, 1992; Kononenko, 1994);

CFS (Hall, 2000); Network-based feature selection

approach (Netzer et al, 2012.); Redundancy Based

Filter (RBF) (Ammu and Preeja 2013).

Concluding the description of the approaches,

the Rough Sets theory conducts tests with all

possible subsets of input features by checking out

which one has better quality of approximation to the

original set (Pawlak, 1982, Hein and Kroenke,

2010).

Examples of feature selection methods

employing this approach are: Rough Sets Theory

(Hein and Kroenke, 2010; Pawlak, 1982); TRS +

Focus2 (Teruya, 2008); RSARSubsetEval (RSAR)

(Chouchoulas and Shen, 2001).

Note that some feature selection methods employ

more than one approach in order to obtain a more

relevant subset of features, refining the selection.

5 METHODOLOGY

Figure 1 presents the proposal of this work to enrich

semantically the feature selection methods.

Basically, this approach is divided into 11 steps as

follow.

1 – A data set with x features {A1, A2...,Ax}, y

tuples and a predictable class {AS} is the input of the

application.

2 – A combination of feature selection filter

methods chooses the most relevant features of the

data set, {M1, M2, …}. This choosing is based just

on the statistical analysis. The possible number of

combination is given by formula (1), where n is the

amount of available feature selection methods and p

is the used quantity of them.

(1)

3 – A subset of the original features is selected,

each one with respective statistical weight (pm). In

this step, it is possible to set up a threshold for the

feature to be accepted and at least one of the feature

selection methods must rank the features.

4 – From the data set, only the name of features

are taken. Those names are compared to an ontology

domain and to a lexical ontology. The used ontology

domain is National Cancer Institute Ontology (NCI,

2014) and the WordNet is a lexical ontology

(Fellbaum,1998; Miller, 1995). Lexical ontology is

used on Thesaurus or dictionaries to recognize

words.

5 – According to the relations between the

features and the predictable class, the names of

features are transformed on concepts from ontology

domain by the use of natural language processing

techniques. If some feature is not found on the

ontology domain, an automatic search for synonyms,

hyperonymy and other relations occurs on WordNet

(Fellbaum,1998; Miller, 1995).

This proposal performs an automatic

normalization procedure with the names of features,

which is the treatment of strings. This procedure

converts strings to lowercase, discards grammatical

accents, deletes blank spaces and hyphens,

withdrawal of numeric digits and punctuation. After

normalization of the strings, the comparison between

feature and an ontology concept is made by

calculating the similarity measure of words (Jaro,

1989; Jaro, 1995; Winkler, 1990) shown in formula

(6), and withdrawing Euzenat Shvaiko (2007).

(2)

(3)

IC3K2014-DoctoralConsortium

(4)

If only if

(5)

(6)

Letters s and t represent two strings to be

compared. Expression com(s,t) represents the

amount of characters that appears in the two strings,

but in a different order. Expression transp (s,t) refers

to the quantity of transpositions of characters

occurred. First calculating σ(s,t) refers to σ

Jaro

(s,t)

calculation. Second σ(s,t) calculation refers to the

measure of σJaro-Winkler (s,t), where P refers to the

size of the common prefix of two strings, and Q is a

constant.

6 – Just the features related to an equivalent

concept are pre – selected as semantically relevant.

Features related to a repeated concepts are removed

because this indicates redundancy. Initially, if a

features is not connected to a concept, it is

considered semantically irrelevant. Then a set of

important concepts, {C1, C2, C3, C7, C8,... CS}, is

selected and will be used on the next steps.

7 – From a base of medical documents, a set of

documents {d1, d2, d3… dn} that contains the

concept equivalent to predictable class (CS) are

recovered. The search looks into abstract, title and

keyword of the documents.

8 – Each concept came from the step 6 is

searching into documents came from step 7.

9 – Each concept k receives a semantic weight

(psk). This semantic weight is calculated by the

formula (8), which uses a modified and weighted

TF-IDF (Term Frequency Inverse Document

Frequency) (Salton e Buckley,1988). Final semantic

weight of each concept is the sum of semantic

weight into each field of the documents: abstract,

title and keyword.

(7)

(8)

10 – Each feature receives its respective

semantic weight, accordingly to related concept.

11 – Subsets came from steps 3 and 10 are faced.

One of three situations can occur:

 If a feature is semantically and statistically

relevant, it receives the normalized sum of

weights (ps and pm);

Figure 1: Architecture of semantic enrichment of feature selection methods.

SemanticEnrichmentofRelevantFeatureSelectionMethodsforDataMininginOncology

 If a feature is just statistically relevant, receives

the normalized statistical weight (pm/||pm||);

 If a feature is just semantically relevant,

receives the normalized semantic weight

(ps/||ps||).

The output of this work is a subset of the most

relevant features.

To understand the formula (8), an example of

calculus of the semantic weight (psk[abstract]) for

the concept Ck was presented in formula (7),

considering the abstract of all selected documents.

Cited calculus has the follow stages:

 TF-IDFck,di is calculated by the multiplication

of TFck,di by IDFck,di, where TFck,di is a

frequency of the concept Ck on document i;

and IDFck,di is the natural logarithm of the

number n of all selected documents divided by

the number of documents that contain the

concept Ck plus 1;

 As the same way, the TF-IDFcs,di would be

calculated, but all selected documents contain

the concept CS related to the predictable

feature. So, just the TFcs,di is important to find

the weighted average psk[abstract];

 Analogously, the weighted average for the

other fields of each document is calculated, in

this case psk [title] and psk[keywords].;

 Final semantic weight of the concept Ck is the

sum of the three partial semantic weights.

Each subset of weights is considered a vector. So,

Euclidean Norm was used as presented in formulas

(9) and (10). It is used to normalize semantic weight

(ps) and statistical weight (pm).

(9)

(10)

6 EXPECTED OUTCOME

In theory, use of semantic enrichment feature

selection methods will bring benefits such as: a)

reduction of the required time to produce mining

models more coherent in the area of cancer and

tumors; b) facilitating of the construction of models

for data mining, since a data miner, without much

knowledge of the physician or genetic field, can

produce good mining models from the semantic

selection of features.

A prototype of the proposal was tested with five

data set related to patients with cancer or tumor as

follow.

Lymphograhy contains 18 features from 148

patients. Prediction feature is the diagnosis if a

patient is normal, has metastasis, has a malignant

lymphoma or a fibrosis.

Breast Cancer has 9 features and the goal is to

classify if there is risk of recurrence in patients who

have already received treatment for tumor.

Location of Primary Tumor presents 17

features of 339 patients. Prediction feature is to

know where primary tumor appeared.

Lymphograhy, Breast Cancer and Location of

Primary Tumor were obtained from the University

Medical Center, Institute of Oncology, Ljubljana,

Yugoslavia and available in http://archive.

ics.uci.edu/ml/. Thanks to M. Zwitter and M. Soklic

for providing this information.

Authorization for Hospitalization from

Brazilian Public Health System (AIH – SUS) - This

data set was obtained from the SIH files available on

site http://www2.datasus.gov.br/SIHD/, option

"Reduced the AIH" menu. It contains 30 attributes

related to hospitalization of 120325 patients with

brain tumors. Prediction feature is the period of stay

of patients, classifying it as short, medium, long or

very long.

Central Nervous System (CNS) is available at

http://www-genome.wi.mit.edu/mpr/CNS/. It con-

tains data of 7129 genes from 60 patients who had

cancer of the central nervous system and had

treatment. Prediction feature is to know if a patient

had survival or not, according to his genetic

characteristics.

Figure 2: Comparative graph for each approach and data

set.

These five data sets were submitted to 6 approaches

of feature selection methods and the results were

compared with the literature concerning Oncology.

Literature came from different sources, from

separate source of the medical repository used to

retrieve documents in this research proposal.

IC3K2014-DoctoralConsortium

Features pointed out in the literature was considered

the gold pattern and F-score calculus was performed

for each approach. Figure 2 presents a comparative

graph.

F-score shows how much the selection of

features by an approach is close to literature. In

other words, how much a feature selection method is

according to human expert and semantic election.

The choosing of some approaches in specifics

data sets is very meaningless, according literature.

For example, this occurs on CFS with Breast Cancer

and on RSAR with Lymphograpy, because there is

no point for them.

For Central Nervous System data set, only the

proposal of this work, named MSSAR, had points.

MSSAR had points in all tested data sets. More

details concerning the tests are available at https://

drive.google.com/folderview?id=0B1CL3P-wkasBT

VBCZFB0UDFadncandusp=sharing.

REFERENCES

Almuallim, H.; Dietterich T. G. 1991. Learning with

Many Irrelevant Features. In: Proceedings of the 9th

National Conference on Artificial Intelligence,

Anaheim, CA, v. 2, pp. 547–552.

Almuallim, H.; Dietterich, T. G. 1992. Efficient algorithms

for identifying relevant features. In: Proceedings of the

Ninth Canadian Conference on Artificial Intelligence,

Vancouver, BC: Morgan Kaufmann. May 11-15, pp.

38-45.

Ammu, P. K.; Preeja, V. 2013. Review on Feature

Selection Techniques of DNA Microarray Data. In:

International Journal of Computer Applications 0975

– 8887 Volume 61– No.12, January 2013. pp. 39-44.

Bray, F.; Ren, J. S.; Masuyer, E.; Ferlay, J. Estimates of

global Cancer prevalence for 27 Sites in the Adult

Population in 2008. Int J Cancer. 2013 Mar 1; 132

(5):1133-45. doi:10.1002/ijc.27711. Epub 2012 Jul 26.

Chahkandi,Vahid; Yaghoobi, Mahdi; Veisi, Gelareh. 2013.

Feature Selection with Chaotic Hybrid Artificial Bee

ColonyAlgorithm based on Fuzzy CHABCF In:

Journal of Soft Computing and Applications. pp. 1-8

Chouchoulas, A.; Shen, Q. 2001. Rough set-aided keyword

reduction for text categorization. Applied Artificial

Intelligence: An International Journal. 159:843-873.

Cover, T. M.; Thomas, J. A. 1991. Elements of Information

Print ISBN 0-471-06259-6 Online ISBN 0-471-20061-

1. 563 p.

Dash, M.; Liu, H. 2003. Consistency-Based Search in

Feature Selection. Artificial Intelligence. 1511-2:155-

176, December, 2003.

Deisy, C., Baskar, S., Ramraj, N., Saravanan Koori, J., and

Jeevanandam, P. 2010.. A novel information theoretic-

interact algorithm (IT-IN) for feature selection using

three machine learning algorithms. Expert Systems

with Applications, 37(12), 7589-7597. Elsevier Ltd.

doi:10.1016/j.eswa.2010.04.084

Euzenat, Jérôme and Shvaiko, Pavel. 2007. Ontology

matching, Springer-Verlag, 978-3-540-49611-3.

Fellbaum, Christiane. 1998. WordNet: An Electronic

Lexical Database. Cambridge, MA: MIT Press.

Ferlay, J.; Soerjomataram, I.; Ervik, M. R.; Dikshit, S.;

Eser, C.; Mathers, M.; Rebelo, M.; Parkin, D.;

Forman, D.; Bray, F. GLOBOCAN 2012 v1.0, Cancer

Incidence and Mortality Worldwide: IARC Cancer

Base No. 11 [Internet]. Lyon, France: Inter-national

Agency for Research on Cancer; 2013. Available at:

http://globocan.iarc.fr, Accessed on June 2014.

Freitas, A. A. 2001. Understanding the Crucial Role of

Attribute Interaction in Data Mining. Artificial

Intelligence Review, (1991), 177-199.

Hall, M. A. 2000. A correlation-based feature selection for

discrete and numeric class machine learning.

ICML'00. In: Proceedings of the 17th International

Conference on Machine Learning. pp. 1157-1182.

He, X.; Cai, D.; Niyogi, P. 2005. Laplacian score for

feature selection. In: Y. Weiss, B. Scholkopf, and J.

Platt, editors, Advances in Neural Information

Processing Systems 18, Cambridge, MA, MIT Press.

Hein, N.; Kroenke, A. 2010 Escólios sobre a Teoria dos

Conjuntos Aproximados – Commentaries about the

Rough Sets Theory. In: Revista CIATEC – UPF, vol.2

1, pp. 13-20. doi: 10.5335/ciatec.v2i1.876 13.

Inbarani, H. H.; Thangavel, K.; Pethalakshmi, A. 2007.

Rough Set Based Feature Selection for Web Usage

Mining. In: International Conf. on Computational

Intelligence and Multimedia Applications. ICCIMA

2007, pp. 33-38. IEEE. doi:10.1109/ICCIMA.2007.356

Jaro, M. A. 1989. Advances in record linkage methodology

as applied to the 1985 census of Tampa Florida.

Journal of the American Statistical Association 84

(406): 414–20. doi:10.1080/01621459.1989.10478785.

Jaro, M. A. (1995). Probabilistic linkage of large public

health data file. Statistics in Medicine. 14 (5–7): 491–

8. doi:10.1002/sim.4780140510. PMID 7792443.

Jemal, A.; Bray, F.; Center, M.; Ferlay, J.; Ward, E.;

Forman., D. 2011. Global Cancer statistics. CA

Cancer Journal for Clinicians.; 61(2):69–90.

Kira, K.; Rendell, L. A. 1992. The Feature Selection

Problem: Traditional Methods and a New Algorithm,

In: Proceedings of 10th Conference on Artificial

Intelligence, Menlo Park, CA, pp. 129-136.

Kira, K.; Rendell, L. A.1992. A practical approach to

feature selection. In: Sleeman and P. Edwards,editors,

Proceedings of the 9th International Conference on

Machine Learning ICML-92, Morgan Kaufmann, pp.

249-256.

Kononenko, I. 1994. Estimating attributes: Analysis and

extension of RELIEFF. In: F. Bergadano and L. de

Raedt, editors, In: Proceedings of the European

Conference on Machine Learning, April 6-8, Catania,

Italy, Berlin: Springer-Verlag, pp. 171-182.

Kuo, Y-T.; Lonie, A.; Sonenberg, L. Domain Ontology

Driven Data Mining: A Medical Case Study.

Proceddings of 2007 ACM SIGKDD Workshop on

SemanticEnrichmentofRelevantFeatureSelectionMethodsforDataMininginOncology

Domain Driven Data Mining (DDDM2007); 2007.

Aug 12-14; San Jose, California, USA, pp.11-17.

Lee, Huei Diana. Seleção de Atributos Importantes para a

Extração de Conhecimento de Bases de Dados. Tese

de Doutorado. USP, 2005. 154p.

Liu, H.; Setiono, R. 1996. A Probabilistic Approach to

Feature Selection: a Filter Solution. In: Proceedings of

the 13th International Conference on Machine

Learning, Morgan Kaufmann. pp. 319–327

Liu, H.; Setiono, R.1998. Feature Selection for Large

Sized Databases. In Proceedings of the 4th World Con-

gress on Expert System, Morgan Kaufmann, pp. 68–75.

Mansingh, G.; Osei-Bryson, K.-M.; Reichgelt, H. 2011.

Using ontologies to facilitate post-processing of

association rules by domain experts. Information

Sciences, 1813, Elsevier Inc. pp. 419-434.

doi:10.1016/j.ins.2010.09.027.

Microsoft 2014. [Online]. Available at: <http://

msdn.microsoft.com/pt-br/library/ms175382.aspx>.

Accessed in 2014.

Miller, George A. 1995. WordNet: A Lexical Database for

English. Communications of the ACM Vol. 38, No.11:

39-41

National Cancer Institute (NCI) [Online]. Available at:

http://www.cancer.gov/ . Accessed in 2014.

Netzer, M.; Fang, X.; Handler, M.; Baumgartner, C. 2012.

A coupled two step network-based approach to

identify genes associated with breast cancer. Proc. 4th

Int. Conf. on Bioinformatics, Biocomputational Sys-

tems and Biotechnologies, (Biotechno, 2012), pp. 1-5.

Osl, M.; Dreiseitl, S.; Cerqueira, F.; Netzer, M.; Pfeifer,

B.; Baumgartner, C. 2009. Demoting redundant

features to improve the discriminatory ability in

cancer data. Journal of Biomedical Informatics,

424,Elsevier Inc. pp. 721-725. doi:10.1016/

j.jbi.2009.05.006

Pawlak, Z. 1982. Rough sets. In: International Journal of

Computer and Information Sciences, vol. 11, New

York, NY. n.º5, pp. 341-356, Plenum. http://roughsets.

home.pl/www/.

Pearson, K. 1901. On Lines and Planes of Closest Fit to

Systems of Points in Space. Philosophical Magazine 2

11: 559–572.

Peng, H.; Long, F.; Ding, C. 2005. Feature Selection

Based on Mutual Information: Criteria of Max-

Dependency, Max-Relevance, and Min-Redundancy.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 278: pp. 1226-1238.

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

approaches in automatic text retrieval. Information

Processing and Management 24 (5): 513–523.

doi:10.1016/0306-4573(88)90021-0..

Tan, K. C.; Teoh, E. J.; Yu, Q.; Goh, K. C. 2009. A hybrid

evolutionary algorithm for attribute selection in data

mining. Expert Systems with Applications, 364, pp.

8616-8630. doi:10.1016/j.eswa.2008.10.013

Tan, P.-N.; Steinbach, M.; Kumar, V. 2005. Introduction to

Data Mining, 1st Edition. Addison-Wesley Longman

Publishing Co., Inc., Boston, MA, USA.

Teruya, Anderson. 2008 Uma metodologia para seleção de

atributos no processo de extração de conhecimento de

base de dados baseada em teoria de rough sets.

Dissertação de Mestrado. Universidade Federal Mato

Grosso do Sul, 86p.

Winkler, W. E. 1990. String Comparator Metrics and

Enhanced Decision Rules in the Fellegi-Sunter Model

of Record Linkage. In: Proceedings of the Section on

Survey Research Methods (American Statistical

Association): 354–359.

Wu, C.-A., Lin, W.-Y., Jiang, C.-L., and Wu, C.-C. 2011.

Toward intelligent data warehouse mining: An

ontology-integrated approach for multi-dimensional

association mining. Expert Systems with Applications,

38(9), 11011-11023. Elsevier Ltd. doi:10.1016/j.eswa.

2011.02.144.

Yu, L.; Liu, H. 2003. Feature selection for high-

dimensional data: A fast correlation-based filter

solution. In: T. Fawcett and N. Mishra, editors,

Proceedings of the 20th International Conference on

Machine Learning ICML-03, August 21-24, Washing-

ton, D.C., 2003. Morgan Kaufmann, pp. 856-863.

Zaki, M.; Meira Jr, W.2009 Fundamentals of Data Mining

Algorithms, Cambridge University Press in press.

555p. Available at: http://www.dcc.ufmg.br/mining

algorithms/

Zhao, Z.; Liu, H.2007. Searching for Interacting Features.

In: Proceedings of the 20th International Joint

Conference on AI IJCAI, January 2007.

Zhao, Z.; Liu, H. 2007. Spectral feature selection for

supervised and unsupervised learning. In International

Conference on Machine Learning ICML, 2007.

IC3K2014-DoctoralConsortium