Abstract:
Feature selection (FS) is a crucial preprocessing step in Data Mining, aimed at enhancing classification performance by
identifying the most relevant features. While numerous techniques for FS exist in the literature, there remains a continuous need to
develop novel methods to achieve superior results. This research article introduces a novel framework designed to form clusters of
features based on user choice and Symmetrical Uncertainty (SU). The framework creates 'N' clusters, from which one dominant
cluster is selected based on the performance of a Multi-Layer Perceptron (MLP) applied to each cluster. Each cluster contains a
unique set of features. The dominant cluster's features are then evaluated using Jrip, J48, and K-Nearest Neighbour (KNN)
classification algorithms, combined with ensembling methods like bagging and boosting. In the proposed methodology features are
grouped in 2 clusters, 3 clusters and 4 clusters. Additionally, the features identified by the proposed method are compared against
those derived from traditional filter-based techniques. The proposed method demonstrates superior performance in most cases. The
effectiveness of this method is validated using a well-known dataset comprising 60 features, highlighting its potential to outperform
conventional FS techniques. This innovative approach addresses the ongoing demand for effective FS methods, contributing to
improved classification accuracy and efficiency in data mining tasks