Feature selection in threes: Neighborhood relevancy, redundancy, and granularity interactivity

Abstract

As a fundamental granular computing strategy, neighborhood granulation has been acknowledged as an intuitive and effective approach to feature evaluation and selection. However, such an approach always has a bias towards a fixed neighborhood granularity, while ignoring the observations across different levels of granularity. To this end, a novel algorithm based on Neighborhood relevancY, redundancY, and granularity interactivitY (N3Y) is proposed. Technically, N3Y adheres well to the rudiment of three-way decision, evaluating and selecting features in threes: 1) feature-to-class relevancy; 2) feature-to-feature redundancy; 3) granularity-to-granularity interactivity. Specifically, firstly, the neighborhood symmetrical uncertainty induced by neighborhood measures is adopted to evaluate the relevancy and redundancy of candidate feature subset; secondly, the proposed neighborhood granularity interactivity allows an uncertainty quantification for finer-to-coarser granularity, and is leveraged as a supplemental factor to guide the relevancy and redundancy, making our procedure more comprehensive; thirdly, a forward-greedy selector is devised, which is required to maximize the evaluation criterion integrating neighborhood relevancy, redundancy, and granularity interactivity. Extensive experiments demonstrate that N3Y outperforms several other advanced feature selectors.