Abstract
Rapidly determining the chemical composition of marine resources is essential for their optimal utilization. Spectroscopy data analysis combined with regression modelling provides an efficient way for estimating chemical properties. Feature selection enhances these models by identifying spectroscopic features relevant to chemical composition. However, most of the existing feature selection methods address the tasks for different chemical components independently, despite shared patterns among related tasks. The paper proposes a novel evolutionary multitasking-based feature selection method to simultaneously address multiple related feature selection tasks. The method integrates a task clustering method to group related tasks, a design method to find an effective uniform preprocessing pipeline for the related tasks, and a multi-objective EMT algorithm to optimize feature subsets jointly. 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.
• Introduced a task correlation-based clustering method to group related feature selection tasks in fish chemical analysis, enabling effective knowledge transfer through evolutionary multitasking (EMT).
• Developed a uniform preprocessing strategy to ensure a common feature space across tasks, which is crucial for the success of multitask learning.
• Proposed an evolutionary multitasking approach for multi-objective feature selection, enabling knowledge sharing across related tasks and improving regression performance.
• Demonstrated through experiments that the proposed framework enhances prediction accuracy in fish chemical analysis while using fewer features.