Improving classification accuracy by means of the sliding window method in consistency-based feature selection

Abstract

In the digital era, collecting relevant information of a technological process has become increasingly cheaper and easier. However, due to the huge available amount of data, supervised classification is one of the most challenging tasks within the artificial intelligence field. Feature selection solves this problem by removing irrelevant and redundant features from data. In this paper we propose a new feature selection algorithm called Swcfs, which works well in high-dimensional and noisy data. Swcfs can detect noisy features by leveraging the sliding window method over the set of consecutive features ranked according to their non-linear correlation with the class feature. The metric Swcfs uses to evaluate sets of features, with respect to their relevance to the class label, is the bayesian risk, which represents the theoretical upper error bound of deterministic classification. Experiments reveal Swcfs is more accurate than most of the state-of-the-art feature selection algorithms.