Abstract
Graph Neural Networks (GNNs) are highly effective for social recommendation, but they often stumble because friends don't always share the same tastes—a problem known as social inconsistency. This paper introduces ConsisRec, a novel GNN framework that dynamically samples and aggregates only the consistent neighbors for a specific user-item pair, effectively filtering out irrelevant social noise and significantly improving rating prediction accuracy.
Introduction
Recommender systems frequently suffer from the "cold-start" problem, where limited interaction data makes it difficult to recommend items to new users. To alleviate this, platforms often incorporate social networks as side information, operating under the assumption that users are influenced by their friends. Recently, Graph Neural Networks (GNNs) have emerged as a powerful tool for this task, as they can simultaneously aggregate information from both user-item interaction graphs and user-user social graphs.
However, existing GNN models typically assume all social connections are equally informative. This paper identifies a critical flaw in this assumption: social inconsistency. A social connection does not guarantee preference alignment for every item. This inconsistency manifests at two levels:
- Context-level: Friends may interact with disjoint item categories (e.g., one friend prefers books while another prefers sports gear).
- Relation-level: Friends may interact with the same item but hold opposing opinions (e.g., divergent ratings for the same movie).
Standard GNNs that aggregate all neighbors equally risk diluting user embeddings with noisy, inconsistent information.
Methodology: The ConsisRec Framework
To address social inconsistency, we propose ConsisRec. Unlike standard GNNs that perform uniform message passing, ConsisRec selectively filters and weighs neighbors through three primary components.
1. Query Layer
When predicting a rating for a user $u$ and item $t$, the model generates a specific "query embedding" by concatenating the embeddings of $u$ and $t$. This allows the model to dynamically identify which friends are relevant to the specific item context.
2. Consistent Neighbor Sampling
The model computes a consistency score between the query embedding and the embeddings of the user's neighbors. It then probabilistically samples a subset of neighbors, favoring those with higher consistency scores. This "hard sampling" approach effectively prunes irrelevant social noise before aggregation.
3. Relation Attention
Following sampling, ConsisRec applies a self-attention mechanism that considers the type of relation (e.g., social link vs. specific rating levels). This handles relation-level inconsistency by assigning higher weights to the most relevant interaction types.
Experimental Evaluation
ConsisRec was evaluated on two real-world datasets: Ciao (7,317 users, 104,975 items) and Epinions (18,069 users, 261,246 items). Both datasets include explicit ratings and social trust links.
Performance Results
ConsisRec was compared against six baselines, including traditional matrix factorization (SoRec, SocialMF) and modern GNN approaches (GraphRec, GCMC+SN). ConsisRec achieved state-of-the-art performance on both datasets:
- Ciao: RMSE of 0.9722 and MAE of 0.7394.
- Epinions: RMSE of 1.0495 and MAE of 0.8046.
The model showed an average relative improvement of 1.7% over the strongest baselines, a significant margin in the context of recommendation accuracy.
Discussion and Analysis
A key insight from the ablation studies is that hard sampling beats soft weighting. Completely dropping inconsistent neighbors via probabilistic sampling was more effective at reducing noise than simply assigning low attention weights.
As shown in the sensitivity analysis, the model is sensitive to the Neighbor Percent hyperparameter. If the sampling ratio approaches 1.0 (aggregating all neighbors), performance degrades significantly as inconsistent neighbors are reintroduced. This confirms that filtering out "noisy" friends is the primary driver of the model's success.
Conclusion
ConsisRec demonstrates that social graphs are inherently noisy for recommendation tasks and that forcing GNNs to aggregate all neighbors can be counterproductive. By making neighbor sampling dependent on the specific item context, ConsisRec effectively filters social inconsistency. Future work may extend this concept of "graph inconsistency" to other domains, such as knowledge graphs or biological networks.
Summary updated