Work

Evolutionary multitasking for multi-objective feature selection in fish chemical analysis

Year: 2025

Type: article

Abstract: Rapidly determining the quality and composition of marine resources is essential for their optimal utilization. Spectroscopy data analysis, along with constructing regression models to relate structured spectral data to the reference data, is an effective and efficient way to quickly measure the chemical properties of marine resources. Performing feature selection tasks allows models to be validated, by confirming that spectroscopic features are related to the chemical data they are predicting. In some cases feature selection can identify features inversely correlated to chemistry of interest. However, most of the existing works address the feature selection tasks for different chemical components independently, even though these tasks are related and share similar spectral patterns. In this paper, a novel feature selection method based on evolutionary multitasking is developed to simultaneouslyaddress multiple related feature selection tasks with the same original set of features on the spectral data, leveraging shared knowledge to enhance accuracy and efficiency. In this framework, a task clustering method is developed to ensure that feature selection tasks with inherent connections are solved together. A preprocessing pipeline design method is developed to find a commonly effective preprocessing pipeline for the related tasks. Furthermore, an evolutionary multitasking multi-objective method is developed to address the feature selection tasks for the chemical components simultaneously. In this evolutionary multitasking method, a search method that effectively selects and shares useful solutions across tasks is developed to achieve positive transfer across tasks. Experiments on feature selection for a fish dataset consisting of 21 regression tasks show that the proposed framework significantly improves the regression performance by sharing knowledge across tasks.

Source: Applied Soft Computing

DOI: 10.1016/j.asoc.2025.114511