Novel Combined Variable Selection Approach Using Memetic Algorithm With Complex Harmonic Regularization

Variable selection has been highly successful in big data analyses, and regularization approaches are commonly used methods, which can automatically select important variables while constructing machine learning models. Due to the real datasets have complex relationships between relevant variable groups, many group-sparsity regularization approaches are proposed recently. However, these approaches usually are non-convex and sensitive to hyper-parameters. Therefore, optimization of the regularizations is a challenging task. In this paper, we present a novel memetic algorithm with the complex harmonic regularization (MA-CHR), which combines EC algorithm for hyper-parameters (global search) and complex harmonic regularization with path seeking strategy for the group-sparsity variable selection (local search). We further introduced a novel genetic individual representation (intron+exon) to efficient obtain the global optimal solution of this group-sparsity regularization. Simulation and five real data experiments demonstrate that the proposed MA-CHR method performs better than the state-of-the-art regularization methods in selecting groups of relevant variables and classification.