Collaborative activity recognition via check-in history

With the growing number of smartphones and increasing interest of location-based social network, check-in becomes more and more popular. Check-in means a user has visited a location, e.g., a Point of Interest (POI). The category of the POI implies the activities which can be conducted. In this paper, we are trying to discover the categories of the POIs in which users are being located (i.e., activities) based on GPS reading, time, user identification and other contextual information. However, in the real world, a single user's data is often insufficient for training individual activity recognition model due to limited check-ins each day. Thus we study how to collaboratively use similar users' check-in histories to train Conditional Random Fields (CRF) to provide better activity recognition for each user. We leverage k-Nearest Neighbors (kNN) and Hierarchical Agglomerative Clustering (HAC) for clustering similar users and learn a separated CRF for each cluster on the histories of its users. As for similarity, the first metric involves linear combination of three types of user factors attained by matrix decomposition on User-Activity, User-Temporal and User-Transition matrices. The second metric between two clusters can be the cosine similarity between weights of CRF corresponding to these two clusters. By the initial experiment on real world check-in data from Dianping, we show that it is possible to improve the classifier performance through collaboration and that the first similarity metric is not good to find the real neighbors.