Fusion of Personalized Implicit Relations for Social Recommendation

In recent years, research in social recommendation has shown that modeling implicit relations effectively alleviates data sparsity and cold-start problems. However, related research often overlooks the influence of user and item rating biases when constructing implicit relations. The divergence in user rating systems and item evaluation standards often results in ratings displaying starkly contrasting extremes, making it difficult to identify implicit relations among users (or items) to fulfill the requirements of personalized recommendations. To this end, we propose a fusion of personalized implicit relations for social recommendation (FIR-REC) based on graph neural networks. First, we regard the average ratings of users and items as benchmarks to eliminate the bias of user-item ratings. Based on this benchmark, user-user (item-item) pairs with the same preferences are regarded as implicit candidate pairs. Considering the heterogeneity of social relationships, we propose a preference hedging formula to calculate preference correlation coefficients for each implicit candidate pair and select the top-k ranked implicit friends for each node to construct a personalized implicit network. Specifically, this formula can utilize different scoring strategies to calculate the preference scores for consistency and inconsistency in implicit pairs. To enhance robustness, it introduces a balancing factor to mitigate the influence of less interacted implicit pairs. Next, we utilize a Graph Attention Network to aggregate neighbor node information in explicit and implicit social relations. Finally, we utilize a user-specific gating mechanism to integrate user representations from explicit and implicit social relations. This helps assess the importance of the two types of relations for different users, enabling precise and stable predictions. Extensive experiments on three open datasets demonstrate the superiority of our model compared with state-of-the-art social recommendation models.