Real-time POI recommendation via modeling long- and short-term user preferences

Abstract

Recently, Next Point-of-Interest (POI) Recommendation which proposes users for their next visiting locations, has gained increasing attention. A timely and accurate next POI recommendation can improve users’ efficient experiences. However, most existing methods typically focus on the sequential influence, but neglect the user’s real-time preference changing over time. In some scenarios, users may need a real-time POI recommendation, for example, when using Take-away Applications, users need recommending the appropriate restaurants at the specific moment. Hence, how to mine users’ patterns of life and their current preferences becomes an essential issue for the real-time POI recommendation. To address the issues above, we propose a real-time preference mining model (RTPM) which is based on LSTM to recommend the next POI with time restrictions. Specifically, RTPM mines users’ real-time preferences from long-term and short-term preferences in a uniform framework. For the long-term preferences, we mine the periodic trends of users’ behaviors between weeks to better reflect users’ patterns of life. While for the short-term preferences, trainable time transition vectors which represent the public preferences in corresponding time slots, are introduced to model users’ current time preferences influenced by the public. At the stage of recommendation, we design a category filter to filter out the POIs whose categories are unpopular in corresponding time slots to reduce the search space and make recommendation fit current time slot better. Note that RTPM does not utilize users’ attributes and their current locations for recommendation, which makes great contributions to users’ privacy protection. Extensive experiments on two real-world datasets demonstrate that RTPM outperforms the state-of-the-art models on Recall and NDCG.