Semantic activity recognition using mobile phone sensors -

Abstract

Recommender systems use contextual and non-contextual information about the user in order to make good and appropriate recommendations. User's emotion is one of the factors that play an important role in these recommendations and is affected by multiple elements such as the user's current activity. Consequently, knowing the user's current activity is essential for making appropriate recommendations. In this thesis, we address the problem of semantic activity recognition using data collected from the user's mobile phone. Our approach recognizes a large set of activities that are comprehensive enough to cover most activities users engage in. Moreover, multiple environments are supported, for instance, home, work, and outdoors. Our approach suggests a multi-level classification model that is accurate in terms of classification accuracy, comprehensive in terms of the large number of activities it covers, and applicable in the sense that it can be used in real settings. Hence, in literature, these three properties are not existent altogether in a single approach. Proposed approaches normally optimize their models for either one or at max two of the following properties: accuracy, comprehensiveness and applicability. When compared to the state-of-the-art in activity recognition from mobile phones, our approach outperforms the state-of-the-art on the fronts of activities types and quantity, environments and settings covered, comprehensiveness, and applicability. We were also able to achieve comparable results in terms of accuracy, while having a significantly higher number of activities.