No abstract

No abstract

Stress has become an important health problem, but existing stress detectors are inconvenient in long-term real-life use because users either have to wear dedicated devices or expend notable interaction efforts in system adaptation to specifics of each person. Adaptation is necessary because individuals significantly differ in their perception of stress and stress responses, but typical adaptation employs supervised learning methods and hence requires fairly large sets of labelled data (i.e. information on whether each reporting period was stressful or not) from every user. To address these problems, we propose a novel unsupervised stress detector, based on using a smartphone as the only device and using discrete hidden Markov models (HMM) with maximum posterior marginal (MPM) decisions for analysis of phone data. Our detector requires neither additional hardware nor data labelling and hence is truly unobtrusive and suitable for lifelong use. Its accuracy was evaluated using two real-life datasets: in the first case, adaptation was based on very short (a few days) phone interaction histories of each individual, and in the second case-on longer histories. In these tests, the proposed HMM-MPM achieved 59 and 70% accuracies, respectively, which is comparable with results of fully supervised methods, reported by other works.

The majority of recommender systems require explicit user interaction (ranking of movies and TV programmes and/or their metadata, such as genres, actors etc), which requires user time and effort. Furthermore, such ranking is often done separately by each person, while merging these manually acquired individual preferences in multi-user environments remains largely an unsolved problem. This work presents a method for learning a joint model of a multi-user environment from implicit interactions: programme choices which family members make together and separately. The proposed method allows to adapt to the practices of each particular family and to protect family privacy, because the joint family model is learned for each family separately. Furthermore, since the accuracy of machine learning methods is family-dependent and none of the machine learning methods outperforms others for all families, a fairly lightweight classifier ensemble selection approach is applied for better adaptation to the specifics of each family. In tests on the real-life TV viewing histories of 20 families, acquired over 5 months, the classifier ensemble achieved an accuracy comparable with that of systems which require explicit user ratings: an average recall of 57% at an average precision of 30%, despite only a few programme metadata descriptors being available.

No abstract