Source-linked AI summary
Graph Neural Networks for Social Recommendation
Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, Dawei Yin
TL;DR
Social recommendation needs to combine user-user relations with user-item interactions, including opinions and varying social strengths. GraphRec jointly models these graphs and signals, and experiments on two real-world datasets show that it outperforms state-of-the-art baselines.
Problem
GNN-based social recommendation must combine user-user and user-item graphs while jointly modeling item opinions and heterogeneous social-relation strengths.
Method
GraphRec is a graph neural network framework that coherently models both graphs and jointly captures interactions, opinions, and heterogeneous social strengths.
Results
GraphRec outperforms state-of-the-art baselines on two real-world datasets.
Takeaways & Limitations
Opinion information plays a crucial role in improving GraphRec’s performance, while the model differentiates social ties by considering heterogeneous strengths.
Takeaways & Limitations
The study uses static rating and social information and does not incorporate rich user or item attributes.
Abstract
from arXiv · showhide
In recent years, Graph Neural Networks (GNNs), which can naturally integrate node information and topological structure, have been demonstrated to be powerful in learning on graph data. These advantages of GNNs provide great potential to advance social recommendation since data in social recommender systems can be represented as user-user social graph and user-item graph; and learning latent factors of users and items is the key. However, building social recommender systems based on GNNs faces challenges. For example, the user-item graph encodes both interactions and their associated opinions; social relations have heterogeneous strengths; users involve in two graphs (e.g., the user-user social graph and the user-item graph). To address the three aforementioned challenges simultaneously, in this paper, we present a novel graph neural network framework (GraphRec) for social recommendations. In particular, we provide a principled approach to jointly capture interactions and opinions in the user-item graph and propose the framework GraphRec, which coherently models two graphs and heterogeneous strengths. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed framework GraphRec. Our code is available at \url{https://github.com/wenqifan03/GraphRec-WWW19}
1 INTRODUCTION
Social recommendation can exploit user-user relations alongside user-item interactions, while GNNs offer a natural way to integrate graph structure and node information. GraphRec addresses the challenges of combining both graphs, modeling opinions, and handling heterogeneous social strengths.
- Social relations can help users filter information and improve recommendation performance.
- GNNs learn representations by transforming and aggregating information from local graph neighborhoods.They integrate node information with graph topology.
- Social recommendation represents users and items through a user-item graph and users through a user-user social graph.Users participate in both graphs and bridge them for latent-factor learning.
- The framework must combine both graphs, jointly capture interactions and opinions, and model heterogeneous social-relation strengths.The user-item graph records both interactions and users’ opinions on items.
- GraphRec coherently models social-recommendation graph data while jointly capturing interactions and opinions and mathematically considering heterogeneous social strengths.The paper reports experiments on real-world datasets to demonstrate the framework’s effectiveness.
2 THE PROPOSED FRAMEWORK
GraphRec models social recommendation by learning user and item representations from both the social graph and user-item graph. It combines opinion-aware interactions with attention-based aggregation to represent heterogeneous influence across users and interactions.
- Framework overview: GraphRec contains user modeling, item modeling, and rating prediction components.User modeling learns latent user factors, item modeling learns latent item factors, and rating prediction uses these representations.
- User modeling: User representations combine item-space information from user-item interactions with social-space information from the user-user graph.The framework uses separate item and social aggregations before combining the resulting latent factors.
- User modeling: GraphRec represents each user-item interaction by fusing an item embedding with an opinion embedding through an MLP.The resulting opinion-aware interaction representation captures both the interaction and the user’s rating-based opinion.
- User modeling: Attention-based item aggregation assigns individualized weights to interactions instead of assuming that every interaction contributes equally.The attention network uses the interaction representation and the target user’s embedding, and Softmax normalizes the resulting scores.
- User modeling: Social aggregation uses attention to model heterogeneous tie strengths and emphasize users that are important for influencing the target user.The social attention relates neighboring users’ representations to the target user embedding.
- Item modeling: Item modeling fuses each interacting user’s embedding with that user’s opinion before aggregating these opinion-aware representations for the item.This allows different opinions about the same item to contribute differently to its latent representation.
3.1 Experimental Settings
The evaluation uses two social-networking datasets, standard rating-prediction metrics, representative baselines, and varied train-validation-test splits and hyperparameters.
- Datasets: The experiments use the Ciao and Epinions1 datasets, where users rate items, write reviews, and add friends to a Circle of Trust.Ratings range from 1 to 5, and opinion embeddings are initialized for the five rating scores.
- Evaluation Metrics: Performance is evaluated with Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), where smaller values indicate better predictive accuracy.Even small MAE or RMSE improvements can affect top-few recommendation quality.
- Baselines: GraphRec is compared with traditional recommender systems, traditional social recommender systems, and deep neural network-based recommender systems.The baselines include PMF, SoRec, SoReg, SocialMF, TrustMF, NeuMF, DeepSoR, and GCMC+SN.
- Parameter Settings: Training uses 80% or 60% of each dataset, with the remainder split equally between validation and testing.Embedding sizes are searched over 8, 16, 32, 64, 128, and 256, alongside several batch sizes and learning rates.
3.2 Performance Comparison of Recommender Systems
GraphRec consistently outperforms the representative baselines on rating prediction across the Ciao and Epinions datasets. The comparisons also indicate benefits from social information and neural network architectures.
- Overall Results: GraphRec consistently outperforms all baseline methods on the Ciao and Epinions rating-prediction comparisons.The comparison reports overall RMSE and MAE performance.
- Social Information: Social recommendation methods using both rating and social-network information outperform PMF, which uses rating information alone.These results support social-network information as complementary to rating information.
- Neural Architectures: NeuMF performs much better than PMF despite both using rating information בלבד, indicating stronger performance from the neural network architecture.
- Neural Social Recommendation: DeepSoR and GCMC+SN outperform the matrix-factorization social recommendation baselines.GCMC+SN is particularly strong among the baselines, consistent with the value of GNN-based graph representation learning.
- Summary: The overall comparison suggests that social information helps recommendation, neural models improve performance, and GraphRec outperforms representative baselines.
3.3 Model Analysis
GraphRec is analyzed through component ablations and hyperparameter studies on Ciao and Epinions. The results assess social information, opinion information, attention mechanisms, and embedding size.
- Social Network Information: Removing social-network information worsens GraphRec’s recommendation performance, indicating that social information contributes to user latent-factor learning.
- Opinions in Interaction: Removing opinion information reduces performance by 3.50% and 2.64% on RMSE and 5.84% and 5.02% on MAE for Ciao and Epinions, respectively.
- Attention Mechanisms: GraphRec-α and GraphRec-µ perform worse than GraphRec, supporting attention mechanisms for item and user aggregation.
- Attention Mechanisms: Removing social attention β significantly reduces performance, indicating that different social friends have different influence during social aggregation.
- Embedding Size: Performance generally increases and then decreases with embedding size: increasing from 8 to 64 helps, whereas size 256 degrades performance.The study identifies a trade-off between representation power and model complexity when selecting embedding length.
4 RELATED WORK
Related work covers social recommendation, neural recommendation models, and GNNs. GraphRec addresses the limited prior attention to applying GNNs specifically to social recommendation.
- Social Recommendation: Social recommendation methods commonly assume that users’ preferences resemble or are influenced by nearby people through social correlations.
- Deep Neural Networks for Recommendation: Deep neural networks have been applied to recommendation using modalities such as audio, text, and images, as well as nonlinear user-item interactions.
- Neural Social Recommendation: NSCR applies neural networks to cross-domain social recommendation but requires users to have one or more social-network accounts.
- Neural Social Recommendation: DLMF and DeepSoR use deep neural networks with ratings, trust relations, user interests, and community effects for recommendation.
- Graph Neural Networks: Existing GNN recommendation methods learn user and item representations from graph data, while relatively little prior work applies GNNs to social recommendation.The paper positions GraphRec as addressing this gap.
5 CONCLUSION AND FUTURE WORK
The conclusion presents GraphRec as a social-recommendation model that jointly uses interactions, opinions, and heterogeneous social tie strengths. Future work targets richer attributes and dynamic rating and social information.
- GraphRec jointly captures interactions and opinions in the user-item graph for social-recommendation rating prediction.
- GraphRec differentiates social tie strengths through heterogeneous-strength modeling, and experiments on two real-world datasets show it outperforms state-of-the-art baselines.
- Future Work: The current model incorporates the social graph but omits other user and item attributes, motivating attribute-aware graph neural networks as future work.
- Future Work: Because rating and social information are treated as static, the paper identifies dynamic graph neural networks for social recommendation as another future direction.