TOWARDS INTERPRETABLE AND INDIVIDUALISED PRECISION DAIRY FARMING: MACHINE LEARNING APPROACHES TO ADVANCE MILK PRODUCTION MODELLING IN HOLSTEIN FRIESIAN COWS

The ability to analyse and model future productivity dynamics in dairy cows is critical for effective herd management, resource allocation, and long-term sustainability in dairy farming. However, understanding and forecasting future milk productivity remains a challenge due to the complex interactions between genetic, environmental, and managerial factors. This doctoral research advances precision dairy farming by leveraging Machine Learning (ML) techniques to develop interpretable and individualised tools for assessing milk productivity in Holstein Friesian cows within Automatic Milking Systems (AMSs). Rather than predicting precise milk yields, this study introduces a novel framework for classifying cows into Productivity Groups (PGs) and investigating the dynamics of milk production across lactation periods. The research was conducted in two phases. In the first phase, a Multi-Clustering framework was developed to define Low and High PGs, integrating results from four distinct clustering algorithms using a novel merging index. The framework was validated across 16 farms, providing insights into herd-level productivity trends and identifying key factors influencing productivity continuity. In the second phase, supervised ML models were employed to automate PG classification and explore the dynamics of future PGs. Additionally, a novel complexity metric was proposed to enhance the interpretability of Genetic Programming (GP) models. Feature importance analysis identified milking robot rate, milking frequency, and milk composition (fat, protein, and lactose percentages) as key predictors of productivity. Interpretable GP models also identified these features, further revealing how they relate with cows' productivity levels. Lastly, this study explored the feasibility of predicting PGs up to two lactation periods ahead. Results indicated that the ability to model future productivity dynamics varied according to the lactation period being forecast, with the highest predictive stability observed for the second lactation and the lowest for the third, highlighting the challenges associated with productivity transitions. The findings contribute to the broader adoption of ML in precision dairy farming by offering robust, interpretable, and actionable insights for herd management.