Source-linked AI summary
ConsisRec: Enhancing GNN for Social Recommendation via Consistent Neighbor Aggregation
Liangwei Yang, Zhiwei Liu, Yingtong Dou, Jing Ma, Philip S. Yu
TL;DR
Social recommendation uses social links alongside user-item interactions for cold-start rating prediction, but existing GNN methods often overlook social inconsistency. ConsisRec selects neighbors using consistency scores and applies relation attention to weight consistent relations, achieving the best results on two real-world datasets with a 1.7% average relative improvement over the second-best method.
Problem
Existing GNN-based social recommendation models often ignore social inconsistency, although social links are not necessarily consistent with rating prediction.
Method
ConsisRec generates item-dependent query embeddings, samples neighbors according to consistency scores, and uses relation attention to emphasize consistent relations.
Results
ConsisRec achieves the best results on the Ciao and Epinions datasets, with a 1.7% average relative improvement over the second-best method.
Takeaways & Limitations
The experiments show the effectiveness of tackling social inconsistency through consistent neighbor selection and relation weighting.
Abstract
from arXiv · showhide
Social recommendation aims to fuse social links with user-item interactions to alleviate the cold-start problem for rating prediction. Recent developments of Graph Neural Networks (GNNs) motivate endeavors to design GNN-based social recommendation frameworks to aggregate both social and user-item interaction information simultaneously. However, most existing methods neglect the social inconsistency problem, which intuitively suggests that social links are not necessarily consistent with the rating prediction process. Social inconsistency can be observed from both context-level and relation-level. Therefore, we intend to empower the GNN model with the ability to tackle the social inconsistency problem. We propose to sample consistent neighbors by relating sampling probability with consistency scores between neighbors. Besides, we employ the relation attention mechanism to assign consistent relations with high importance factors for aggregation. Experiments on two real-world datasets verify the model effectiveness.
1 INTRODUCTION
Social recommendation fuses social links with user-item interactions to address cold-start rating prediction, but existing GNN methods often overlook social inconsistency. ConsisRec addresses this by selecting consistent neighbors and weighting consistent relations during aggregation.
- Social recommendation incorporates social information with user-item interactions to alleviate the cold-start problem.Social networks can influence users’ online behaviors, motivating their use as side information for recommendation.
- GNN-based social recommendation models aggregate information from social and user-item graphs simultaneously.These models rely on neighbor aggregation to learn node representations for recommendation.
- Social inconsistency means social links are not necessarily consistent with the rating prediction process.It includes context-level discrepancies between connected users’ item contexts.
- ConsisRec samples neighbors using consistency scores between query and neighbor embeddings, then applies relation attention to emphasize consistent relations.The framework learns node embeddings for rating prediction from consistent contexts and relations.
- Experiments on two real-world datasets show ConsisRec’s effectiveness.The paper identifies this as a contribution alongside its consistent node-embedding framework.
2 PRELIMINARIES
Social recommendation combines incomplete user-item ratings with a user-user social graph to complete the rating matrix. The problem is represented as a heterogeneous graph containing users, items, and multiple relation types.
- The social recommendation problem contains users, items, an incomplete rating matrix, and a user-user social graph.Ratings encode user preferences, while social edges represent connections such as trust.
- Its objective is to complete the rating matrix by fusing rating information with the social graph.The formulation uses a heterogeneous graph containing both user and item nodes.
- The heterogeneous graph distinguishes user-item links by rating scores and includes social and item-item relations.On Ciao, six rating values produce eight relation types in total.
3 PROPOSED MODEL
ConsisRec addresses social inconsistency by conditioning neighbor selection on the user-item query and weighting sampled neighbors by relation importance. Its framework combines embeddings, query-based sampling, relation attention, multilayer propagation, rating prediction, and RMSE optimization.
- Framework overview: ConsisRec combines an embedding layer, query layer, neighbor sampling, and relation attention within a GNN framework.The model can use multiple propagation layers, although Figure 1 illustrates a one-layer version.
- Embedding layer: Trainable node embeddings represent users and items, while relation embeddings characterize relation-level social inconsistency.
- Query layer: A query layer maps concatenated user and item embeddings to select neighbors consistently for each user-item pair.The query is dynamic because a user may consult different friends for different items.
- Neighbor sampling: Neighbor sampling assigns probabilities from consistency scores, emphasizing neighbors whose hidden embeddings are closer to the query.The sampling ratio γ controls how many neighbors are selected relative to a node’s total neighbors.
- Relation attention: Relation attention assigns importance factors to sampled neighbors by incorporating the relations associated with their edges.The attention weights are computed with a self-attention mechanism using relation embeddings.
- Rating prediction and optimization: After L-layer propagation, ConsisRec predicts ratings with the user-item embedding inner product and trains using RMSE over rating edges.Adam optimization uses a weight decay rate of 0.0001 to avoid over-fitting.
4.1 Experimental Setup
The evaluation uses Ciao and Epinions social-recommendation datasets, six baseline methods, MAE and RMSE metrics, and predefined train-validation-test splits with grid-searched hyperparameters.
- Datasets: Ciao contains 7,317 users, 104,975 items, and 111,781 social links, while Epinions contains 18,069 users, 261,246 items, and 355,530 social links.
- Baselines: ConsisRec is compared with six baselines spanning matrix factorization, collaborative graph embedding, and GNN-based methods.The baselines are SoRec, SocialMF, SoReg, CUNE, GCMC+SN, and GraphRec.
- Metrics: MAE and RMSE evaluate rating prediction, with lower values indicating better performance.Small improvements in both metrics may affect top-N recommendation quality.
- Protocol: Each dataset is split randomly into 60% training, 20% validation, and 20% testing data.Hyperparameters are tuned by grid search, and one GNN layer is used for both datasets.
4.2 Performance Evaluation
ConsisRec achieves the best results on both Ciao and Epinions, outperforming the second-best method by 1.7% on average.
- Overall results: 1.7% relative improvement over the second-best method is achieved on average across Ciao and Epinions.
- Overall results: ConsisRec achieves the best results on both Ciao and Epinions.GCMC, GraphRec, CUNE, and ConsisRec outperform SoRec, SoReg, and SocialMF.
4.3 Ablation Study
The ablation study evaluates the query layer, neighbor sampling, and relation attention, finding that the complete ConsisRec model performs best and that each component contributes to performance.
- Ablation Study: Variant A removes the query layer, which selects neighbors using user embeddings instead of query embeddings.
- Ablation Study: Variant B removes neighbor sampling and aggregates all neighbors, whereas variant C removes relation attention and assigns equal weights to different relations.
- Ablation Study: ConsisRec consistently achieves the best performance against variants A, B, and C.
- Ablation Study: Removing neighbor sampling in variant B dramatically spoils performance, supporting the importance of selecting consistent neighbors.
- Ablation Study: The worse performance of variants A and C demonstrates the importance of the query layer and relation attention, respectively.
4.4 Parameter Sensitivity
Parameter sensitivity analysis on Ciao examines neighbor percent, embedding size, and learning rate, showing that excessive neighbor aggregation harms performance and that tuning matters.
- Neighbor percent: Error increases obviously when neighbor percent rises from 0.8 to 1.0, resulting from aggregating inconsistent neighbors.
- Embedding size: The best embedding size on Ciao is 16.Smaller sizes insufficiently represent node information, while larger sizes lead to over-fitting.
- Learning rate: Learning rate has a critical impact on model performance and needs careful tuning.
5 CONCLUSION AND FUTURE WORK
ConsisRec identifies social inconsistency in social recommendation and modifies GNNs to address it. Experiments on two real-world datasets show its effectiveness, while neighbor filtering and broader graph applications remain future work.
- ConsisRec contains three GNN modifications designed to tackle social inconsistency in social recommendation.
- Experiments on two real-world datasets show ConsisRec's effectiveness.
- Future work includes better filtering of informative neighbors and identifying inconsistency problems in other graph-related research directions.