Abstract:
Feature interaction is a vital aspect of Machine Learning (ML) algorithms, and gaining a deep understanding of these interactions can significantly enhance model performance. This paper introduces the BukaGini algorithm, an innovative and robust approach for feature interaction analysis that capitalizes on the Gini impurity index. By exploiting the unique properties of the BukaGini index, our proposed algorithm effectively captures both linear and nonlinear feature interactions, providing a richer and more comprehensive representation of the underlying data. We thoroughly evaluate the BukaGini algorithm against traditional Gini index-based methods on various real-world datasets. These datasets include the High School Students' Performance (HSSP) dataset, which examines factors affecting student performance; Cancer Data, which focuses on identifying cancer types based on gene expression; Spambase, which targets spam email classification; and the UNSW-NB15 dataset, which addresses network intrusion detection. Our experimental results demonstrate that the BukaGini algorithm consistently outperforms traditional Gini index-based methods in terms of accuracy. Across the tested datasets, the BukaGini algorithm achieves improvements ranging from 0.32% to 2.50%, underscoring its effectiveness in handling diverse data types and problem domains. This performance gain highlights the potential of the BukaGini algorithm as a valuable tool for feature interaction analysis in various ML applications.
