GEFS-AO: A novel graph-based evolutionary feature selection method from the view of auxiliary optimization

Abstract

Graph-based evolutionary feature selection (FS) algorithms attract much attention, since they can simultaneously utilize the advantages of graph and evolutionary computation for solving FS problem. Despite that, due to the encoding of evolutionary algorithms, they often need to search in a complex and large feature graph space. To this end, these algorithms develop different evolutionary operators to overcome the challenge. Unlike the existing algorithms that directly deal with complex optimization problem, this paper tackles the problem from the perspective of auxiliary optimization, where the original complex optimization problem is decomposed into two simple yet complementary auxiliary optimization subproblems. Specifically, in the proposed algorithm (GEFS-AO), a fully connected feature graph is firstly constructed, from which two simple auxiliary feature subgraphs are created. One subgraph only uses some important nodes (features) and all their edges, which ensures the latent important feature combinations could be detected. Another subgraph uses all the nodes and each node only has one edge, which guarantees all features are considered. On each subgraph, a population is evolved to obtain its feature subsets, which is viewed as one auxiliary optimization subproblem. During the evolution, a pair of information exchanging strategies is designed between two auxiliary optimizations, which can adjust the structures of two auxiliary subproblems and improve the performance of both auxiliary optimization subproblems. Moreover, a node weight update strategy is also suggested for two auxiliary subgraphs, which further enhances the quality of final feature subsets. Experimental results on different FS datasets demonstrate the effectiveness of the proposed GEFS-AO.