Multi-Criterion Feature Selection Based on Clustering Symbolic Regression

Feature selection is a highly effective technique for processing high-dimensional datasets. However, existing methods often prioritize the accuracy of constructed models as the main criterion, leading incomplete and unstable outcomes. In this study, the multi-criterion feature selection algorithm based on clustering symbolic regression (FSSR) is proposed to tackle the issue of simplistic criteria in feature selection. The involvement and sensitivity of features are considered as new criteria of feature selection in FSSR. These criteria can be calculated based on the occurrence frequency and partial derivatives in the expressions. Moreover, Pareto non-dominated sorting is incorporated to fully account for both criteria. Furthermore, a clustering symbolic regression algorithm is proposed to preserve local information during the process. By clustering the training set into subsets, the generalization of expressions is improved to enhance the accuracy of feature selection. When applied to the creep life dataset of high-temperature nickel-based alloys, a new FSSR algorithm with domain-specific prior knowledge is proposed for producing more accurate and interpretable results in feature selection