Chi-MIC-share: a new feature selection algorithm for quantitative structure–activity relationship models

Abstract

Quantitative structure–activity relationship models are used in toxicology to predict the effects of organic compounds on aquatic organisms. Common filter feature selection methods use correlation statistics to rank features, but this approach considers only the correlation between a single feature and the response variable and does not take into account feature redundancy. Although the minimal redundancy maximal relevance approach considers the redundancy among features, direct removal of the redundant features may result in loss of prediction accuracy, and cross-validation of training sets to select an optimal subset of features is time-consuming. In this paper, we describe the development of a feature selection method, Chi-MIC-share, which can terminate feature selection automatically and is based on an improved maximal information coefficient and a redundant allocation strategy. We validated Chi-MIC-share using three environmental toxicology datasets and a support vector regression model. The results show that Chi-MIC-share is more accurate than other feature selection methods. We also performed a significance test on the model and analyzed the single-factor effects of the reserved descriptors.