Feature Selection Using Neighborhood Positive Region Certainty

Source abstract

Neighborhood rough set–based attribute reduction is a powerful tool used widely in areas such as machine learning, pattern recognition, and decision support to handle numerical data. Before performing a classification task, it is necessary to find a subset of features that possesses the same classification ability as the entire feature set. To address this requirement, numerous neighborhood rough set–based attribute reduction algorithms have been developed and applied to numerical data. These algorithms almost exclusively utilize positive region information to assess the classification ability of attributes, with minimal reliance on boundary region information. This study proposes a new efficient reduction algorithm using neighborhood positive region certainty (NPRC). It fully leverages both positive region and boundary region information, leading to a significant enhancement of algorithm performance. Firstly, we introduce a novel technique termed neighborhood partition, aiming to gain a deeper understanding of neighborhoods and reveal valuable knowledge. Subsequently, we develop a new model called the partitioned neighborhood rough set, which revolutionizes the rules for determining the region to which an object belongs. Furthermore, we put forward an attribute evaluation method, referred to as NPRC. It not only considers positive region objects but also takes into account the contribution of the boundary region objects to the positive region, extending its value from 0 or 1 to a continuous value between 0 and 1. This innovation provides a more concrete and comprehensive description of the classification ability of attributes. Finally, we design a new attribute reduction algorithm that utilizes NPRC to evaluate attributes and guide a greedy search process to find an optimal subset of features. Experimental results demonstrate that the proposed algorithm is capable of discovering a smaller number of attributes and achieves better classification performance compared to other available algorithms.