A Mixed Effects Machine Learning Framework for Dairy Cattle Methane Prediction

Due to the difficulty of recording dairy cattle (DC) methane (CH4) emissions, as well as the scale of data required for robust generalisation, DC CH4 emission prediction models are commonly trained on datasets compiled from multiple experiments, involving various treatments, and different sets of cattle. These underlying sources of variation, also known as random effects (REs), such as the treatment/management in each specific experiment, or the genetics of each individual cow, can introduce biases into the model, and produce misleading results when the models trained upon them are applied in alternative scenarios. Therefore, we developed a Mixed Effects Machine Learning framework (MEML), which could incorporate biological, environmental and genetic data, to produce refined machine learning (ML) models that could address this issue. The framework initially makes predictions using a ML model, which are then passed into a linear mixed effects (ME) model as an offset, where the residual variation is partitioned between the REs within a dataset. These RE adjustments are then used to correct the original response and a new ML model is retrained on its corrected version, further isolating the remaining residual variation to redistribute between the REs within the dataset, with the process repeating until convergence. This allows the refined ML models produced through the MEML framework to gain a greater appreciation of the authentic relationships between the features and response, apathetic to the RE influence embedded within them, improving their generalisation to external datasets with inevitable differences in RE influences.