Abstract
Efficient herd management requires identifying which cows are consistently high or low performers over time. While Automatic Milking Systems (AMSs) collect rich data on milk yield and cow behavior, translating these into actionable insights for characterizing productivity remains challenging, especially in a way that is both automated and human-understandable. In this study, we present an interpretable machine learning framework to automatically classify cows in AMSs into low and high Productivity Groups (PGs), focusing on Decision Trees and Multi-Objective Genetic Programming (MOGP). The models were trained using a comprehensive dataset derived from AMSs, including production metrics and behavioral indicators. They aimed to distinguish between Low and High PGs, previously defined through a robust unsupervised multi-algorithm clustering approach applied to the same dataset and features. Specifically, the supervised models were tasked with learning to replicate the resulting classification boundaries. In addition to the interpretable models, eXtreme Gradient Boosting (XGBoost) and Support Vector Machines were included to benchmark predictive performance. To deepen the understanding of both the data and the models, we conducted feature importance analyses using model-intrinsic metrics and model-agnostic techniques, including ReliefF and Shapley values. Milking Robot Rate and Milking Frequency consistently emerged as the most critical features. MOGP yielded transparent mathematical expressions, enabling structural insights on the results of feature importance, and enhancing interpretability on the milking productivity levels distinguished by the PGs previously defined by unsupervised machine learning algorithms. Our findings suggest that interpretable models can classify PGs with competitive accuracy while offering practical insights into milk productivity of AMS cows, thereby potentially supporting trust and utility in data-driven and model-driven dairy herd management.