A GA-like Dynamic Probability Method With Mutual Information for Feature Selection

Abstract

Feature selection plays a vital role in promoting the classifier's performance. However, current methods ineffectively distinguish the complex interaction in the selected features. To further remove these hidden negative interactions, we propose a GA-like dynamic probability (GADP) method with mutual information which has a two-layer structure. The first layer applies the mutual information method to obtain a primary feature subset. The GA-like dynamic probability algorithm, as the second layer, mines more supportive features based on the former candidate features. Essentially, the GA-like method is one of the population-based algorithms so its work mechanism is similar to the GA. Different from the popular works which frequently focus on improving GA's operators for enhancing the search ability and lowering the converge time, we boldly abandon GA's operators and employ the dynamic probability that relies on the performance of each chromosome to determine feature selection in the new generation. The dynamic probability mechanism significantly reduces the parameter number in GA that making it easy to use. As each gene's probability is independent, the chromosome variety in GADP is more notable than in traditional GA, which ensures GADP has a wider search space and selects relevant features more effectively and accurately. To verify our method's superiority, we evaluate our method under multiple conditions on 15 datasets. The results demonstrate the outperformance of the proposed method. Generally, it has the best accuracy. Further, we also compare the proposed model to the popular heuristic methods like POS, FPA, and WOA. Our model still owns advantages over them.