Graph based feature selection using ensemble clustering

Abstract

One crucial method for high-dimensional data is feature selection. The main goal of feature selection is to eliminate features that aren’t relevant. However, eliminating unnecessary features is just as crucial. Based on the concept of ensemble clustering, we suggest a new feature subset selection approach. Our technique builds an entire graph on feature space and divides it using several social network graph partitioning algorithms. Ensemble clustering is used to determine the optimal partitioning, and the most ”representative” feature with the strongest correlation to the target class is chosen from each cluster to create the final feature subset. The algorithm is tested on benchmark data sets with dimensionalities ranging from 8 to 168 features, and classification serves as validation. The outcomes demonstrate how well the suggested method eliminates superfluous and unnecessary elements. Classifier accuracies using the chosen features are comparable to those of the most recent methods suggested in the literature, and the number of features chosen using the suggested method is quite small.