A frequent pattern-based coevolutionary framework for multi-component spectral feature selection

Abstract

Spectral feature selection plays a crucial role in spectral analysis as it aims to identify the most effective features from the original high-dimensional wavelength variables, thereby enhancing the accuracy of concentration prediction models. In multi-component spectral feature selection (MCSFS) problems, diverse composition and concentration of samples result in complex overlapping peaks and correlations among variables. This complexity poses challenges in finding optimal subsets of features efficiently. To address this issue, this paper proposes a frequent pattern-based coevolutionary framework for solving MCSFS problems. Specifically, the algorithm starts by generating a main population for multi-component spectral feature selection and multiple auxiliary populations for single-component spectral feature selection. Furthermore, it introduces a frequent pattern mining strategy to identify dynamic superior feature combinations and their updated weights in each population, dealing with the complexity of variables to accelerate the search for effective features. The proposed coevolutionary framework facilitates interactions between populations by sharing the identified feature combinations and offspring information, leading to the acquisition of high-quality feature selection results. Experimental results on twelve MCSFS problems, based on three high-dimensional spectral datasets, demonstrate that the proposed algorithm outperforms six state-of-the-art evolutionary algorithms.