Correlation Coefficient Based Feature Selection Framework Using Graph Construction

Abstract

In machine learning, selecting the best features for classification is a critical issue. It is anecessary task to reduce the number of attributes/features existed in the initial feature space forachieving the outstanding classification accuracy, to minimize the computing power, and toreduce the memory size. In this present research, a novel methodology is proposed based on theconcept of Symmetrical Uncertainty (SU) and Correlation Coefficient (CCE) by constructingthe graph to select the reduced feature set. The recommended features by the proposedmethodology are clubbed into finite number of groups (clusters) by measuring their CCE andconsidering the highest SU score of the feature. From each group, a feature which has maximumSU value is picked up and rest of the features in the same group are ignored. The proposedstructure was inspected with ten (10) real world data sets available in the public domain.Experimental outcomes guarantees that the proposed method is recorded the better performancethan most of the traditional filter based feature selection methods. The proposed methodperformed better than traditional methods such as Information Gain and Chi-Square on 70 % ofthe data sets. It is also produced better result than traditional Gain ration method on 80 % of thedata sets and competing with traditional ReliefF approach on 50 % of the data sets. Thismethodology is assessed using Lazy, Tree Based, Naive Bayes, and Rule Based learners.