KITHAR: a Knowledge Informed Tree based Neural Network for Interpretable Human Activity Recognition

Human Activity Recognition (HAR) seeks to automatically identify various types of human activities from data collected through different mechanisms. Although existing HAR methods achieve high accuracy, they face challenges in interpretability, particularly in fields requiring classification explanations, such as human-computer interaction and sports science. Inspired by physics-informed neural networks and decision trees, a novel interpretable HAR model named KITHAR is proposed. This model incorporates physical knowledge into the generation process of a neural decision tree, allowing the resulting tree to integrate physical prior knowledge and hence enhance model interpretability. Experimental results reveal that this method significantly improves interpretability at both feature and decision tree levels. Additionally, classification accuracy only decreased by 1% compared to the standard method.