NEURO-SYMBOLIC AI APPROACHES FOR SENSOR-BASED HUMAN ACTIVITY RECOGNITION

Sensor-based Human Activity Recognition (HAR) is an active research area, with relevant applications in healthcare and well-being. Deep Learning (DL) classifiers are currently the leading approach to tackle HAR, but their deployment is often limited by their inherent opacity and the scarcity of labeled training data. Fortunately, common sense and domain knowledge about activity execution can improve purely data-driven approaches. Indeed, in the general machine learning community, Neuro-Symbolic AI (NeSy) methods are emerging to combine DL models with more traditional symbolic AI techniques that rely on knowledge-based reasoning to improve models' interpretability while reducing their reliance on labeled data during training. This thesis explores innovative NeSy solutions proposed to enhance sensor-based HAR. The initial chapters focus on NeSy methods designed to mitigate the scarcity of labeled training data. Considering smart-home environments inhabited by multiple subjects, a main problem is data association, i.e., correctly associating sensor events (e.g., the opening of the fridge) with the subject(s) that actually generated them. Most works in the literature addressed this challenge with purely data-driven solutions, thus aggravating the labeled data scarcity problem. For this reason, we propose a NeSy method that relies on symbolic reasoning to tackle data association without the need for any labeled data. Moreover, we also address data scarcity for context-aware HAR based on mobile devices. While NeSy approaches have been already proposed in this research area, they rely on domain knowledge only after the training process of the DL classifier. This limits its ability to handle data uncertainty. Hence, we present two novel NeSy approaches that infuse domain knowledge into DL classifiers during their learning process. Experimental results show how such methods reduce the amount of labeled data required during training while being more robust to noisy data compared to state-of-the-art NeSy methods. Finally, we present an initial investigation of interpretability aspects. We introduce a metric that quantitatively evaluates, based on domain knowledge, the quality of explanations obtained from DL activity classifiers. Due to time constraints, this metric has been currently used only to evaluate purely data-driven approaches. Nonetheless, we plan to employ such a metric to quantify the interpretability benefits provided by NeSy methods for HAR. Overall, all the methods presented in this thesis have been experimentally evaluated on publicly available datasets that have been collected in controlled or in-the-wild settings.