Joint Spatio-Temporal Modeling and Generative Adversarial Network for Point-of-Interest Recommendation

The rapid advancement of Location-based Social Networks (LBSNs) has been observed, and the recommendation of attractive places through Point of Interest (POI) has become increasingly important. However, as the user and POI numbers continue to increase, recommender systems for POIs face various challenges: (1) The challenge lies in the integration of contextual details such as the spatial and temporal coordinates of points of interest (POIs); (2) Scarcity of check-in data. To tackle these challenges, we propose a unique joint model called GEOGAN that combines Geographical Temporal Modeling (GT) with Generative Adversarial Networks (GAN). This approach aims to learn non-linear relationships between users and POIs while also addressing issues related to sparse data. Recent studies have identified a clustering pattern in human mobility behavior on LBSNs, where individuals tend to visit locations that are grouped closely together. The effectiveness of this spatial clustering phenomenon has been demonstrated in enhancing POI recommendations. Therefore, we integrate geographic data into the conventional matrix factorization model. Furthermore, we utilize GANs to iteratively optimize the model by playing a mini-max game. Specifically, the discriminative model aims to enhance the estimation of POI ranking by leveraging unlabeled data selected by the generative model and guiding the training process of the generative model to account for potential associations between users and POIs. Conversely, as a challenger to the current discriminative model, the generative model generates challenging instances for discrimination objective minimization in an adversarial manner. The proposed approach was validated on two publicly available datasets, and the results confirm its feasibility and effectiveness.