In order to predict and recommend the points of interest (POIs) for a given combination of user, current location, and time, a joint probabilistic generative model called the topic region model (TRM) is presented. It combines user interests and user mobility models that consider the semantic, spatial, and temporal dimensions of previous check-in behavior data. The user interest model captures POI semantics using the latent topic model, where each topic is represented with a distribution of words and a distribution of check-in times. The user mobility model assumes a Gaussian distribution for each region with density of check-in locations. The joint effect of user interest and region is represented with a distribution over POIs in the same region that have a similar semantics. Using Foursquare and Twitter check-in data, the TRM parameters are learned and used to generate the top-k POIs.
In addition, to enable the real-time recommendation of POIs, TRM-online is proposed, where the check-in streams are sampled with particle filter concepts to update the TRM parameters. To make it more efficient, it uses a clustering-based branch and bound algorithm to prune the POI search space for fast computation of top-k POIs. The experiments compare various base lines and other approaches with TRM-batch and TRM-online versions, and the two versions of the TRM joint model show their effectiveness in terms of accuracy and efficiency.
The paper would be more interesting if the authors had shared the learned parameters such as latent topic and word distributions, region-related distributions, as well as some concrete lists of recommended POIs. The readability and interest in this topic are obscured by abstract models, algorithms, and experimental results that may not be reproducible.