Interpretable Machine Learning Techniques for Predictive Cattle Behavior Monitoring

Cattle behavior prediction and monitoring is crucial in determining the health of cattle. This study intends to evaluate various Machine Learning (ML) algorithms, using the data collected from 18 cows. Data collection was done from sensors, which included the Afimilk Silent Herdsman Collar that records the raw acceleration values from a 3-axis MEMs accelerometer at a frequency of 10 Hz, and the Rumiwatch Halter device that measure the pressure created by the jaw movements and provide classifications at 10Hz. The data collected include the classifications of the behaviors viz. "Other behavior", "Ruminating", "Drinking/eating". The data was first pre-processed before analysis. Feature extraction was done on the dataset to create 94 new features. Classification performance was investigated with 8 machine learning models (gradient boosting, random forests (RF), extreme gradient boosting (XGBoost), K-nearest neighbors (KNN), logistic regression, classification and regression trees (CART), support vector machines (SVM), Naive Bayes). The accuracy scores of the models were 0.81, 0.80, 0.80, 0.77, 0.76, 0.75, 0.64, 0.56 respectively. For model interpretability, XAI techniques were applied, specifically LIME (Local Interpretable Model-agnostic Explanations) and SHAP (Shapley Additive Explanations). Additionally Eli5 was used to investigate the feature's importance for the best performing model. Many studies have proposed various ML methods for cattle behavior prediction without interpretatbility. This study aims to interpret the predictions of these ML algorithms.