Social Recommendation Based on Preference Disentangle Aggregation
Zheng Li, Qun Liu, Youmin Zhang · 2021
Graph Neural Networks (GNNs) have further promoted the development of social recommendation, but traditional social aggregation models combined with GNNs usually take the highly entangled and complex representations of nodes. However, due to the diversity of node preferences in social recommendation, the establishment of connectivity relationships between nodes just needs to be based on a part of preferences, so it is unreasonable to consider all the preferences of neighboring nodes when aggregating. To address this issues, we propose a novel social recommendation framework based on preference disentangled aggregation (PDARec). This method first maps the coarse-grained comprehensive representation of nodes to more fine-grained latent preference subspaces, then aggregates the information of neighboring nodes in separate latent preference subspaces. At the same time, to ensure the independence of multiple latent preference subspaces, the model uses a preference regularizer to measure the distance of each space. Experiments on two open datasets shows that the proposed social recommendation method based on preference disentangled aggregation produces a better potential representation of users (items) and effectively improves the performance of the recommendation compared with five baseline models, and the metrics RMSE and MAE had dropped 1%~2%.