Attribute reduction with personalized information granularity of nearest mutual neighbors

Abstract

Neighborhood-based attribute reduction plays a vital role in pattern recognition, for selecting a series of informative and relevant attributes from data sets. The increase in dimensionality and complexity of data along with superfluous and noisy attributes, results in significant and tricky issues when reasoning data. To address the ineffectiveness of classical attribute reduction within the neighborhood framework, a competent attribute reduction scheme based on a novel concept of nearest mutual neighbor-based personalized information granularity is proposed in this paper. This study comprises three phases. First, we intend to integrate the strategies of personalized radius and two-sided neighbors into the modeling of information granularity, and both personalized and mutual neighborhood rough sets are developed. In the second phase, the concept of the belief function from evidence theory is introduced to construct an adaptive neighborhood information granule with the guidance of the principle of justifiable granularity; in this sense, the search for a justifiable granularity is transformed into an optimization issue. In the third phase, a new neighborhood homogeneous evidence-based attribute significance evaluation is discussed to avoid the inadequacy of voting-based attribute measures. In addition, an enhanced forward greedy algorithm is proposed to address non-monotonic circumstances. For the experimental evaluation, fifteen benchmark data sets, involving six high-dimensional microarray data sets, are chosen to display the performance of the method and algorithm. The experimental findings reveal the superiority of our work from different perspectives.