Abstract: Location-based Social Networks (LBSNs), which contain the friend relationships between users and the trajectories that users check in POIs, have been widely used to recommend POIs to users, plan events for users and predict the friend relationships between users. However, existing methods for predicting the friend relationships usually ignore the users’ preferences, spatial constraints and dynamic changes for POIs hidden in the check-in trajectories, which have important effects on friend relationship prediction. To address the above issues, a Friendship Prediction method Integrating Multi-View information (FPIMV) is proposed to enhance the embeddings of user pairs and improve friendship prediction by capturing the spatial-temporal information in the check-in trajectories and the friend relationships between users. Specifically, in the spatial information embedding layer of FPIMV, we propose a method for learning the embeddings of user pairs based on the characteristics of check-in POIs, which effectively captures the users’ preferences and spatial constraints for POIs by learning the similarity and the proximity of POIs. In the temporal information embedding layer of FPIMV, we propose a method for learning the embeddings of user pairs based on the characteristics of check-in times, which effectively captures the users’ dynamic changes for POIs by learning the local and global features of check-in times. In the social relationship embedding layer of FPIMV, we learn the embeddings of user pairs based on the meta-paths, which capture the semantic relationships between users. Finally, we enhance the embeddings of user pairs by fusing the above embeddings learned. Experimental results on three datasets from LBSNs show that compared with the baselines, FPIMV improves the performance of friendship prediction.

Key words: location-based social networks, friendship prediction, network embedding, spatial-temporal information

摘要： 位置社会网络（LBSNs）是一类融合了朋友关系和用户签到轨迹的社会网络，其广泛应用于兴趣点推荐、活动策划和朋友关系预测等任务。然而，现有的朋友关系预测方法在处理用户签到轨迹时，在空间维度上忽略了用户的兴趣点偏好和空间约束，在时间维度上忽略了用户兴趣点偏好的动态变化。针对上述问题，提出融合多视角信息的朋友关系预测方法（FPIMV），旨在通过捕获用户签到轨迹中的空间、时间信息以及用户间的朋友关系等多视角信息，有效学习用户对嵌入，提升朋友关系预测效果。具体来讲，FPIMV模型在签到位置嵌入层，提出基于兴趣点相似性和邻近度的用户对嵌入方法，该方法有效捕获了用户的兴趣点偏好和空间约束；在签到时间嵌入层，提出基于签到时间局部和全局特征的用户对嵌入方法，该方法有效捕获了用户兴趣点偏好的动态变化；在社会关系嵌入层，基于元路径学习用户对嵌入，以捕获用户间的语义关系；最后，融合上述嵌入，提高用户对嵌入质量，进而提升朋友关系预测性能。在三个LBSNs数据集上的实验结果表明，FPIMV模型在朋友关系预测任务上优于基线模型。

关键词: 位置社会网络, 朋友关系预测, 网络嵌入, 时空信息