CPSORCL: A Cooperative Particle Swarm Optimization Method with Random Contrastive Learning for Interactive Feature Selection

Genome-wide association study (GWAS) is an important strategy to analyze the genetic basis of complex diseases. However, although GWAS has achieved great success, it is still difficult to fully understand the complexity of diseases while only considering single feature at each time. Selecting interactive features has become a novel perspective to uncover the genetic mechanism of diseases. In this study, we proposed a cooperative particle swarm optimization method, named CPSORCL, for interactive feature selection. The highlights of CPSORCL are adaptive random contrastive learning strategy, flipping strategy based on feature weight, and deep search strategy. The adaptive random contrastive learning strategy adjusts the topological structure according to the population evolution state, establishes a good competition and cooperation mechanism among particles, and hence maintains the population diversity. The flipping strategy based on feature weight dynamically adjusts the probability of feature flip, which effectively realizes the balance between global search and local detection in the solution space. The deep search strategy accurately searches features in the candidate pool to select final interactive features. Experiments were carried out on simulated data sets and age-related macular degeneration data set, and compared with seven popular methods. The experimental results show that CPSORCL is promising in selecting interactive features, and may become an alternative to existing methods. The source codes are available online at https://github.com/CDMBlab/CPSORCL .