An explainable machine learning for geographical origin traceability of mussels Mytilus edulis based on stable isotope ratio and compositions of C, N, O and H

Abstract

Geographical traceability of marine bivalves has been explored continually, however the traceability results have not been recognized owing to the absence of a reasonable explanation for a certain model prediction. To address this problem, we employed an explainable machine learning to identify the origin of mussels based on stable isotope ratio and compositions of carbon (C), nitrogen (N), oxygen (O), and hydrogen (H). Our findings proved that the Extreme gradient boosting (XGBoost) model may be the best model based upon its high accuracy (93.75 %), precision (93.75 %), recall (94.51 %), F1 score (93.72 %) and AUC (0.990). The feature impact for model out of XGBoost was interpreted by SHapley Additive exPlanations (SHAP) on global and local perspective. This study demonstrated the potential of explainable machine learning to elucidate the complex relationships between desirable geographical information of mussels and the selected features.