Introduction
Social Recommendation (SR) systems leverage user-user social ties alongside traditional user-item interactions, based on the premise that friends often share similar preferences. However, real-world social graphs are frequently sparse and noisy, leading to poor generalization in standard Graph Neural Network (GNN) models. SR-GCA (Social Recommendation via Graph-Level Counterfactual Augmentation) addresses these challenges by generating "what-if" counterfactual social and item links. This approach creates a richer training signal, allowing the model to uncover underlying causal relationships rather than merely memorizing noisy correlations.
Methodology
The SR-GCA framework integrates a standard GNN backbone with a novel Graph-Level Counterfactual Augmentation module.
The architecture consists of four primary components:
- SR Backbone: A lightweight GCN (similar to LightGCN) processes user-item, user-user, and item-item graphs independently before aggregating them into final embeddings.
- Counterfactual Links Constructor (GCA-CLC): This module identifies "what-if" scenarios. For any node pair, it searches the embedding space for "nearest neighbors" with opposite link statuses. For instance, if two users are not friends, the system finds a similar pair who are friends to serve as a counterfactual example.
- Counterfactual Links Learner (GCA-CLL): An MLP-based decoder learns to predict both factual (observed) and counterfactual (simulated) links.
- Joint Optimization: The model is optimized using a combination of BPR loss for recommendation ranking, counterfactual loss for link refinement, and a discrepancy distance regularization term to align factual and counterfactual distributions.
Experimental Setup
The model was evaluated on two benchmark datasets:
- Ciao: 6,672 users, 98,875 items, and 109,503 social ties.
- Epinions: 11,111 users, 190,774 items, and 203,989 social ties. These datasets provide both explicit interaction data and a social network layer, making them ideal for testing SR models.
Key Results and Discussion
SR-GCA consistently outperformed ten state-of-the-art baselines, including DSL and MHCN. On the Ciao dataset, it achieved an 8.72% improvement in Hit Ratio (HR@10) over the closest competitor.
Beyond raw performance, the model demonstrated significant transferability. Adding the GCA module to other models like DGNN and GDMSR resulted in a 2-3% performance boost. Furthermore, SR-GCA showed high robustness; when synthetic noise was added to the social graph, it maintained higher relative accuracy than competitors, suggesting that counterfactual training effectively filters out social noise. Finally, the model is efficient, training 37.5% faster than the SMIN model on the Ciao dataset despite the additional counterfactual computations.
Conclusion and Future Work
SR-GCA demonstrates that graph-level counterfactual reasoning can significantly enhance social recommendation by identifying stable causal features. While effective, the current model treats the user-item bipartite graph with less granularity than social or item-item graphs. Future research could explore more detailed counterfactuals for specific user-item interactions and investigate more sophisticated clustering strategies for counterfactual generation.
Summary updated