A Novel Classification Method Based on Stable Feature Selection for High Dimensional Missing Data

ABSTRACT

High‐dimensional data is now prevalent in AI applications, and stable feature selection methods are commonly used to handle it for improving application accuracy and yielding reasonable feature subsets. However, frequently occurring missing data can weaken conventional methods, reducing both classification performance and feature‐selection stability. To address this problem, we propose stable feature selection based on particle swarm optimization (SFSPSO) for high‐dimensional incomplete data. SFSPSO combines hybrid filter‐based ranking, disturbance particle initialization, evolutionary selection with imputation, and similarity mutation to guide the search, handle missing values, maintain population diversity, and preserve stability. We compare SFSPSO with eight alternative methods on five datasets at missing data rates from 5% to 50%. Results on four classification metrics and one stability metric show that SFSPSO generally achieves better performance while retaining an acceptable overhead.