Personal Context Recognition from Sensors

Machine learning has become one of the most emerging topics in a lot of research areas, such as pervasive and ubiquitous computing. Such computing applications always rely on the supervised learning approach to recognize user’s context before a suitable level of services are provided. However, since more and more users are involved in modern applications, the monitored data cannot be guaranteed to be always true due to wrong information. This may cause the mislabeling in machine learning and so affects the prediction. The goal of this Ph.D. thesis is to improve the data quality and solve the mislabeling problem caused by considering non-expert users. To achieve this goal, we propose a novel algorithm, called Skeptical Learning, aiming at interacting with the users and filtering out anomalies when an invalid input is monitored. This algorithm guarantees the machine to use the pre-known knowledge to check the availability of its own prediction as well as the label provided by the users. This thesis clarifies how we design this algorithm and makes three main contributions: (i.) we study the predictability of human behavior through the notion of personal context; (ii.)we design and develop Skeptical Learning as a paradigm to deal with the unreliability of users when providing non-confidential labels that describe their personal context; (iii.) we introduce an MCS platform where we implement Skeptical Learning on top of it to solve unreliable labels issue. Our evaluations have shown that Skeptical Learning could be widely used in pervasive and ubiquitous computing applications to better understand the quality of the data relying on the machine knowledge, and thus prevent mislabeling problem due to non-expert information.