An Unsupervised Approach for Selection of Candidate Feature Set Using Filter Based Techniques

Abstract

High dimensionality is the one of the important issue in preprocessing stage of data mining.Initial feature space may have irrelevant or redundant features. These properties of featuresdecrease the performance of classifier, and also require more memory and high computingpower. This issue can be addressed by selecting the best feature subset for improving theclassification performance. In this research, we have proposed an unsupervised approach usingfilter based feature selection methods and K-Means clustering technique to derive the candidatesubset. Score of each feature is calculated using traditional filter based methods. Then Min-Maxtechnique is applied to normalize the dataset. K-Means algorithm is employed on the dataset toform the clusters of features. To decide the strong subset, Multi-Layer Perceptron(MLP) isapplied on each cluster. Best cluster is selected based on the minimum Root Mean Square(RMS) error rate given by MLP. This framework is compared with traditional methods over sixwell known datasets having the total features in between 34 and 90 using various classificationalgorithms. The proposed method recorded 75% competitive rate than Information Gain(IG),71% success rate than Gain Ratio Attribute Evaluator(GR) and Chi Square AttributeEvaluator(Chi), 83% competitive rate than ReliefF(Rel) traditional methods. Jrip classifierperformed 55%, J48 recorded 66%, Naive Bayes displayed 88%, IBK (Instance Based)displayed 80% success rate over all the datasets.