Abstract

The ability to acquire and understand context is a key requirement to build intelligent, adaptive, and personalised services and environments. In the last decades we have experienced the impact of spatio-temporal awareness enabled by the localisation technologies of mobile computing. Pervasive computing has a major role in exploiting additional sensing modalities obtaining large streams of sensor data that, processed with AI techniques, can take context awareness to a new level, enabling human-centered applications that could not be considered before. This talk reviews the advances in this topic and the lessons learned. In particular, it considers sensor-based human activity recognition as a task to derive high level context from low level context data and sensors’ signals. The talk identifies open challenges and new research opportunities. Among the challenges there is the scarcity of annotated data as required by deep learning methods, the need to explain the activity prediction process, and the privacy threats involved in releasing personal data. Among the opportunities, some of the limits of purely statistical approaches could be effectively improved by combining them with reasoning on common knowledge in terms of ontologies or knowledge graphs. Data scarcity and privacy concerns may be also mitigated by a combination of techniques including active learning and distributed approaches like federated learning.