Explainable models for predicting crab weight based on genetic programming

Reliable weight prediction plays an important role in commercial transactions and population management of crabs. Existing research usually used predefined models to explain the relationship between the weight and the length of crabs. In this paper, we propose an effective regression method using genetic programming (GP) to build explainable models, which include more features to explore potential relationships between the weight and the physical features of crabs. The GP-based method has been evaluated on a publicly available dataset of crabs. The experimental results were compared with several baseline methods for predicting two kinds of crab weights. GP shows the best performance among all the baseline methods on the test set, i.e., 90.8% for predicting the weight of crabs and 81.3% for predicting the shucked weight of crabs in terms of coefficient of determination. Thanks to the explicit ability of feature selection, GP can select more important features to improve the prediction performance. More importantly, the generated models can provide potential interpretability, which is particularly valuable for domain experts in fisheries management and ecological research. • GP-based symbolic regression method is proposed to predict two types of crab weight. • GP outperforms traditional and ML methods such as ANN and random forest. • GP’s feature selection capability is validated by Sobol analysis and PFI analysis. • Models derived from GP have advantages of interpretability and transparency.