From When to Where: A Multi-task Learning Approach for Next Point-of-Interest Recommendation

Abstract

Temporal information plays a crucial role in analyzing user behaviors in Location-Based Social Networks (LSBNs). Different from existing methods such as Matrix Factorization (MF) and Recurrent Neural Networks (RNNs) methodologies that only make use of historical temporal information, we try to explore the prediction of the user’s future check-in time to help POI recommendation in this paper. We propose a new multi-task neural network recommendation model, namely MTNR, which jointly learns when and where users are likely to go next. To learn the user’s next POI preference based on the next check-in time prediction, we introduce two kinds of time-decay POI transition tensors to calculate the user’s common and personal POI transition probability, respectively. By combining the POI preference learned in POI and time prediction tasks, MTNR can get better recommendation accuracy in POI recommendation. We conducted experiments on three real-world datasets. The result shows that our model significantly outperforms well-known methods.