Multiple GRAphs-oriented Random wAlk (MulGRA2) for Social Link Prediction

Current link prediction methods in Location-Based Social Networks (LBSNs) fuse graphs derived from users' check-in data and their social links to form a single graph or network. Then, they learn node representations and link probabilities from a fused graph that undermines the distinctive characteristics of each user's spatiotemporal mobility and social links. Consequently, the input datasets are heavily contaminated with noise, which makes it challenging for these algorithms to make accurate predictions. Our study employs the proposed Multiple GRAphs-oriented Random wAlk (MulGRA2) to model graphs while maintaining the distinctive characteristics of the data to address the issue of noisy data in the learning process. Specifically, we use three graphs: a social graph constructed from social links data, a user co-occurrence graph derived from users' check-in data to capture spatiotemporal co-occurrence, and a user-location bipartite graph that links users to specific locations based on the same check-in data. After traversing all three graphs, it learns node representation and infers links effectively. Extensive experiments on both Foursquare and synthetic datasets demonstrate that our algorithm significantly improves link prediction performance. Furthermore, sensitivity analysis on various datasets confirms the robustness of our algorithm.