Hierarchical neighborhood entropy based multi-granularity attribute reduction with application to gene prioritization

Abstract

As a prominent model of granular computing, neighborhood rough set provides clear granularity organization and expression in terms of inherent parameter (neighborhood radius). Such characteristic is widely captured in a plenitude of attribute reduction procedures, while igniting a tricky issue of tuning parameters. In this study, we therefore propose a parameter-free multi-granularity attribute reduction scheme. Fundamentally, our scheme applies three-way decision as thinking in threes. First, data-aware multi-granularity structure is automatically induced from self-contained distance space instead of manually edited or appointed granularities. Second, a novel multi-granularity feature evaluation criterion named hierarchical neighborhood entropy is defined to measure the feature significance. Finally, a sequential forward searching algorithm is designed to find the optimal reduct. With application to gene prioritization, our method performed on microarray data is experimentally demonstrated to be more effective and efficient in differentially expressed genes discovery as compared with other well-established attribute reduction algorithms.