LBSN REPRESENTATION LEARNING BASED ON RANDOM WALK STRATEGY

In order to fully mine the cross correlation information and improve the accuracy and robustness of LBSN, a learning method of LBSN representation based on random walk strategy is proposed. Based on the analysis of community overlapping structure, a role decomposition algorithm was designed for the hypergraph. This algorithm described the allocation of user context and role nodes, and thus fully captured the three behaviors in LBSN. The further introduced random walk strategy captured the geographic impact and time cycle of user movement through time decay factor, which could strongly reflect the latest preferences of users and help overcome data sparsity and dynamic preferences. The validity of the proposed method was proved on several datasets.