Bio-Inspired Feature Selection via an Improved Binary Golden Jackal Optimization Algorithm

Feature selection is an efficient way to filter the most valuable subset of features from the high-dimensional feature space of the original data sets, which can remove part of irrelevant and redundant features while keeping the classification accuracy constant or increasing. In this paper, an improved binary golden jackal optimization (IBGJO) which extends from the conventional golden jackal optimization (GJO) is proposed for wrapper-based feature selection. First, we use a fitness function to linearly connect the classification accuracy and the number of selected features. Second, eight kinds of transfer functions divided into two families, S-shaped and V-shaped, are introduced as the binary version of the GJO and evaluated the classification performance. Finally, the IBGJO proposes three improvement mechanisms which are a novel population initialization with logistic chaotic map and opposition-based learning, a local search strategy with feature aggregation and a global search factor with dynamic movement to enhance the performance in solving feature selection problems. Simulation experiments based on 16 classical data sets in the UC Irvine Machine Learning Repository are conducted and the performance of IBGJO is compared with several peer algorithms, and the experimental results illustrate that the proposed IBGJO outperforms other compared algorithms.