Graph representation learning on Location-Based Social Networks

With the popularity of online Social Networks, location-based social networks (LBSN) have accumulated massive data and have been widely used in the research of mining user behavior preferences due to their rich Spatio-Temporal and semantic information. Nevertheless, the traditional manual extraction of LBSN features is limited and time-consuming. In recent years, graph representation learning has been successfully applied to the modeling and representation of various graph structure data such as recommendation systems and knowledge maps, demonstrating its powerful non-linear fitting and representation learning capabilities. However, most of the existing Graph representation learning studies focus on static and homogeneous networks, and it is difficult to simultaneously combine time, location information, and social relationships to capture the complex structure and user preferences in LBSN, which makes it difficult to extract effective information from LBSN. Therefore, this paper proposes a two-stage Graph representation learning framework TGE-LBSN (Two Stages of Graph Embedding on LBSN) for LBSN, which transforms LBSN into a heterogeneous network, and automatically extracts the features of LBSN with the help of graph representation learning to obtain nodes' vector representation with sufficiently rich information and utilize the prediction and recommendation tasks in the social domain to verify its effectiveness. First of all, biased sampling is carried out on the check-in hyperedge of LBSN according to users' Check-in time. In the first stage, the IVGS (Initial Vector Generation Stage) algorithm is designed, the friendship edges and the Check-in super edges are used to jointly generate nodes' vectors containing position and feature information by IVGS algorithm. The generated nodes' vectors are used as the input of the second stage. In addition, the second stage is mainly responsible for generating the final nodes' vectors in LBSN. LBSN is divided into different subgraphs according to the Check-in time, and we design the LBSN-oriented SAN (Select Aggregated Neighbors) strategy which is used to select representative neighbors to complete the aggregation operation, and then use the subgraph vector generation algorithm SVG (Subgraph Vector Generation) to obtain the vector representation of the nodes in each subgraph. Finally, the loss function is set according to the downstream tasks, and the attention mechanism is also used to learn adaptive weights for the subgraphs in different time periods to obtain the final vectors of the nodes, and then we use the final nodes' vectors to complete various prediction tasks in the social domain. Plenty of comparative experiments are carried out with the benchmark methods on the real LBSN data sets and on the time series social network, respectively, we use ROC curve as the evaluation standard. Extensive experimental results verify that the proposed algorithm TGE-LBSN outperforms other benchmark methods, and it can efficiently extract the effective information of LBSN and retain it in the embedding vector of the node. Specifically, in terms of friendship prediction, the AUC value can be increased by up to 42% compared with the existing models. On the point of interest recommendation task, the AUC value can be increased by up to 7% compared with the benchmark algorithm. © 2022, Science Press. All right reserved.