A graph-based sparse filter feature selection method for multi-class and ordinal data

Feature selection is a common pre-processing step for high-dimensional datasets. Filter-based methods rank feature relevance independently of specific classifiers. This work proposes a graph-based filter method for multi-class classification, aiming to drastically reduce features while retaining valuable class-separability information. The approach combines the Jeffries-Matusita distance with diffusion maps for nonlinear dimensionality reduction. Features are eliminated based on their distribution in the low-dimensional space, selecting a small subset with complementary separation strengths. In addition, we propose a modified version to address the case of ordinal feature selection. Experimental results on public tabular and time-series datasets demonstrate the method's effectiveness compared to existing filter-based techniques.