Multi-objective binary grey wolf optimization for feature selection based on guided mutation strategy

Feature selection aims to choose a subset of features with minimal feature-feature correlation and maximum feature-class correlation, which can be considered as a multi-objective problem. Grey wolf optimization mimics the leadership hierarchy and group hunting mechanism of grey wolves in nature. However, it can easily fall into local optimization in multi-objective optimization. To address this, a novel multi-objective binary grey wolf optimization based on a guided mutation strategy (GMS), called MOBGWO-GMS, is proposed. In the initialization phase, the population is initialized based on feature correlation, and features are selected using a uniform operator. The proposed GMS uses the Pearson correlation coefficient to provide direction for local search, improving the local exploration ability of the population. Moreover, a dynamic agitation mechanism is used for perturbation to prevent population stagnation due to the use of a single strategy. The strategy is dynamically adjusted to maintain population diversity and improve detection ability. To evaluate the classification ability of quasi-optimal subsets, a wrapper-based k-nearest neighbor classifier was employed. The effectiveness of the proposed algorithm was demonstrated through an extensive comparison with eight well-known algorithms on fourteen benchmark datasets. Experimental results showed that the proposed approach is superior in the optimal trade-off between the two fitness evaluation criteria and can easily jump out of local optima compared to other algorithms.