Anomaly detection to infer context changes in temporal data

Context-aware recommendation systems (CARS) have a distinct advantage over other types of recommenders in scenarios where contextual information, such as the user’s location, the weather, the presence of holidays, etc., can significantly enhance the relevance and effectiveness of recommendations. Retrieving contextual information for a specific dataset is often not trivial or even possible; the context model and its contextual features are often imposed at design time and are equal for all users, hence the need for a methodology that can infer such information in its absence and in a personalized way. Anomaly detection is a collection of techniques used to identify data points that deviate from the norm or from their expected behaviour. In this paper, we present a novel approach that leverages known anomaly detection techniques based on LSTM to extract a specific type of contextual information (context change) from existing datasets. We implement a tool that detects context change in temporal datasets and tests it against a dataset of Fitbit data collected from two participants in this case study. The results show that the technique easily highlights periods of time when the users’ context was indeed different from their norm, thus successfully detecting a change in their context.