Modified Recursive Feature Elimination Analysis For Best Dataset Feature: A Case Study On Social Assistance Data

Abstract

Evaluating social assistance data plays a crucial role in determining recipient eligibility and optimizing resource allocation. However, a major challenge lies in selecting the most relevant features without clear knowledge of their impact. Recursive Feature Elimination (RFE) is a widely used method for feature selection, but it requires predefined decisions on the number of features to retain—often causing uncertainty among practitioners. This study proposes a novel approach called K-Optimal Recursive Feature Elimination (KRFE) to enhance unsupervised learning in social assistance data evaluation. KRFE modifies the standard RFE by introducing a stopping criterion based on repeated performance-weight evaluations using K-Means and K-Medoid clustering models. The optimal number of features (k) is determined by analyzing feature distance and integrating unsupervised weighting to assign importance scores. A case study was conducted on a complex social aid dataset. The process includes model training, separation value computation, feature ranking, and elbow-point analysis to identify the optimal K. Results show that the optimal feature set is achieved when k = 3, yielding a separation score of 0.8 with K-Means and 1.0 with K-Medoid. Furthermore, KRFE with K-Medoid converged in only 4 iterations compared to 10 with K-Means, indicating improved efficiency. These findings confirm that KRFE can effectively identify the optimal feature subset and that K-Medoid enhances RFE performance in unsupervised contexts involving complex datasets.