Home // DBKDA 2025, The Seventeenth International Conference on Advances in Databases, Knowledge, and Data Applications // View article

Evaluating the Potential of SHAP-Based Feature Selection for Improving Performance

Authors:
Ashis Kumar Mandal
Basabi Chakraborty

Keywords: Feature Selection; SHapley Additive exPlanations (SHAP); classification models; machine learning.

Abstract:
Feature selection is an important preprocessing step in developing efficient and accurate classification models. Among various techniques, recently SHapley Additive exPlanations (SHAP)-based feature selection has gained attention for its interpretability and ability to quantify the contributions of individual features to model predictions. This study investigates the effectiveness of SHAP-based feature selection technique, specifically focusing on Linear SHAP, in improving classification performance. The research utilizes 10 diverse datasets to evaluate Linear SHAP's capability in identifying relevant features for classification tasks. The performance of Linear SHAP is assessed across varying percentages of selected features and compared to classification models without feature selection. Three popular filter-based feature selection approaches: Chi-square($Chi^2$), Mutual Information, and Correlation-based methods are also used for feature selection with the same bench mark data sets. Comparative analysis, supported by statistical significance tests, demonstrates that Linear SHAP performs equally well to the traditional methods while offering the added benefit of interpretability. The findings suggest that Linear SHAP is a viable and promising alternative to established feature selection techniques in the realm of classification tasks.

Pages: 19 to 24

Copyright: Copyright (c) IARIA, 2025

Publication date: March 9, 2025

Published in: conference

ISSN: 2308-4332

ISBN: 978-1-68558-244-9

Location: Lisbon, Portugal

Dates: from March 9, 2025 to March 13, 2025