COLLABORATIVE APPROACHES FOR SENSOR-BASED HUMAN ACTIVITY RECOGNITION IN DATA SCARCITY SCENARIOS

One of the most important goals of Human Activity Recognition (HAR) is to automatically obtain information on the behaviors of the users to proactively assist them with their tasks. In the literature, the majority of physical activity recognition approaches rely on fully- supervised techniques to collaboratively train a recognition model over the data collected from a large number of users. However, these solutions usually suffer from numerous issues like scalability, privacy, poor personalization, and scarcity of labeled training data. In this thesis, we will focus on analyzing in deep those problems, with the scope of proposing novel methodologies to tackle them. First of all, we consider the labeled data scarcity issue. Indeed, obtaining human-annotated activity examples is costly, intrusive, time-consuming, and hence unpractical on a large scale. Semi-supervised approaches have been suggested to reduce the size of the training set required to initialize the model, but their effectiveness revealed not satisfactory for those activities that involve similar body movements (e.g., standing and taking the elevator). In order to mitigate this problem, we propose a novel hybrid semi-supervised and knowledge-based framework that uses the context that surrounds users (e.g. semantic location, speed, weather) to enable a machine learning model trained with a limited number of labeled data to classify a wide set of context-dependent activities. Then, we consider the scalability and privacy issues that arise in collaboratively training a recognition model with the data coming from a large number of different users. Federated Learning (FL) showed to be a promising paradigm to address these problems. However, most of the FL-based solutions for HAR proposed in the literature assume that users can always obtain labeled data to train the recognition model, hence inheriting the limitation related to human annotation that we mentioned before. Moreover, generating a single global model for all the users may not be as effective as expected. Indeed, different subjects could perform activities in different ways depending on their physical traits and habits. In order to tackle these problems, we introduce innovative hybrid semi-supervised and FL-based solutions that enable personalized, privacy-aware, and scalable activity recognition. In conclusion, we analyze the possible information leakage of FL for HAR, with the aim of obtaining hints to guide the future development of specific privacy-preserving techniques.