Cost-sensitive feature selection based on adaptive neighborhood granularity with multi-level confidence

Abstract

Neighborhood rough set model is considered as one of the effective granular computing models in dealing with numerical data. This model is now widely discussed in feature selection and rule learning. However, there is no theoretical analysis on the issue of neighborhood granularity selection, the influence of sampling resolution, test and misclassification costs on modeling. In this paper, we design an adaptive neighborhood rough set model according to data precision and develop a fast backtracking algorithm for neighborhood rough sets based cost-sensitive feature selection by considering the trade-off between test costs and misclassification costs. In the proposed model, the neighborhood granularity, based on the 3σ rule of statistics, is adaptive to data precision that is described by the multi-level confidence of the feature subsets. Our experiments, thoroughly performed on 12 datasets, demonstrate the effectiveness of the model and the efficiency of the backtracking algorithm.