Symbolically Regressing Fish Biomass Spectral Data: A Linear Genetic Programming Method With Tunable Primitives.

Area of Science:

• Chemometrics
• Machine Learning
• Bioinformatics

Background:

• Spectral data analysis is crucial for fish production, providing insights into fish meat's chemical properties.
• Existing machine learning methods struggle with noisy, limited fish biomass spectral data, hindering pattern discovery.
• Symbolic regression offers a potential framework for analyzing complex spectral datasets.

Purpose of the Study:

• To develop an improved machine learning approach for analyzing fish biomass spectral data.
• To enhance the accuracy and interpretability of fish composition prediction models.
• To address challenges posed by noise and limited training data in spectral analysis.

Main Methods:

• Modeled fish biomass spectral data analysis as a symbolic regression problem.
• Employed a linear genetic programming method with newly proposed tunable primitives.
• Tuned inherent coefficients of primitives to improve regression model approximation ability.

Main Results:

• Achieved improved overall performance in fish biomass composition prediction across over ten targets.
• Synthesized compact and interpretable regression models, highlighting key spectral features.
• Demonstrated good generality across various spectral data treatments and other symbolic regression tasks.

Conclusions:

• The proposed linear genetic programming method with tunable primitives effectively analyzes fish biomass spectral data.
• The approach enhances prediction accuracy and interpretability, outperforming existing methods.
• The method shows promise for broader applications in spectral data analysis and symbolic regression.