Interaction-based feature selection using Factorial Design

Feature interaction provides insight into hidden domain knowledge and interactive structure of a data set. In feature selection, identifying significant interactions among features is a challenging task. Since possible candidates of interactions increase exponentially to the number of features. In this paper, we propose a two-stage feature selection approach that makes full use of interactions. In the first stage, we decompose the feature selection problem into a sum of interaction information. Then, higher-order interactions are used to select an interaction-preserving feature subset. In the second stage, we employ design of experiments (DOE) to identify significant interactions from the feature subset. The proposed approach is compared with mRMR, JMIM and ReliefF. Experiments on public available data sets demonstrate that our approach reveals the influence of interactions and so that outperforms the state-of-the-art filter methods.