A Hybrid Improved Multi-objective Particle Swarm Optimization Feature Selection Algorithm for High-Dimensional Small Sample Data

Traditional feature selection methods have great limitations for processing high-dimensional small sample data, and it is difficult to accurately and efficiently propose the optimal feature subset. To improve the feature selection ability to deal with high-dimensional data, this paper proposes a hybrid improved multi-objective particle swarm optimization feature selection algorithm based on the classification accuracy and feature selection number, called HIMOPSO. The algorithm uses the inherent attributes of the data to compute and filter the features quickly and uses multi-objective particle swarm optimization to perform a fine search. The parameters of particle velocity are modified nonlinearly according to the number of iterations. The explosion of new particles creates a pool of candidate particles to select the new particles as the next generation of initial particles. This increases population diversity, allows particles to explore more potential areas, and improves the overall population quality. The final feature subset is classified on the gene expression profile data. Experimental results show that this method is superior to the comparison algorithm in classification accuracy and feature number and can balance these two objectives.

Automated summary

Traditional feature selection methods have limitations in processing high-dimensional small sample data and proposing optimal feature subsets accurately and efficiently. To improve feature selection ability, this paper proposes a hybrid improved multi-objective particle swarm optimization feature selection algorithm, called HIMOPSO, based on classification accuracy and feature selection number. The algorithm quickly computes and filters features using data's inherent attributes and performs a fine search using multi-objective particle swarm optimization. Experimental results demonstrate that this method surpasses comparison algorithms in classification accuracy and feature number, while balancing these two objectives.