Source-linked AI summary
Dual Graph Attention Networks for Deep Latent Representation of Multifaceted Social Effects in Recommender Systems
Qitian Wu, Hengrui Zhang, Xiaofeng Gao, Peng He, Paul Weng, Han Gao, Guihai Chen
TL;DR
Existing social recommendation methods often treat friends’ effects as static and may conflate friendship with preference similarity, despite effects varying across contexts. DANSER uses dual GATs to model four user- and item-domain effects and a contextual-bandit policy to fuse them; experiments on benchmark and commercial datasets report improved recommendation accuracy over state-of-the-art methods.
Problem
Social recommendation seeks to use social information to address data sparsity and cold-start problems, but prior methods often assume friendship implies similar preferences and model social influence statically.
Method
DANSER uses dual graph attention networks for user- and item-domain homophily and influence, plus a contextual multi-armed-bandit policy to dynamically weight their four interactions.
Results
Experiments on one benchmark dataset and one commercial dataset report that DANSER improves recommendation accuracy over state-of-the-art social recommendation methods and that its proposed components are effective.
Takeaways & Limitations
Modeling multifaceted social effects across users and related items provides effective representations for recommendation and supports heterogeneous interaction weights across user-item pairs.
Takeaways & Limitations
The paper identifies prior approaches’ limitations but supplies no explicit limitation of DANSER in the provided passages.
Abstract
from arXiv · showhide
Social recommendation leverages social information to solve data sparsity and cold-start problems in traditional collaborative filtering methods. However, most existing models assume that social effects from friend users are static and under the forms of constant weights or fixed constraints. To relax this strong assumption, in this paper, we propose dual graph attention networks to collaboratively learn representations for two-fold social effects, where one is modeled by a user-specific attention weight and the other is modeled by a dynamic and context-aware attention weight. We also extend the social effects in user domain to item domain, so that information from related items can be leveraged to further alleviate the data sparsity problem. Furthermore, considering that different social effects in two domains could interact with each other and jointly influence user preferences for items, we propose a new policy-based fusion strategy based on contextual multi-armed bandit to weigh interactions of various social effects. Experiments on one benchmark dataset and a commercial dataset verify the efficacy of the key components in our model. The results show that our model achieves great improvement for recommendation accuracy compared with other state-of-the-art social recommendation methods.
1 INTRODUCTION
The paper identifies static, preference-similarity assumptions as inadequate for multifaceted and context-dependent social effects, then proposes DANSER to model four effects across users and items and fuse them dynamically.
- Motivations and Rationales: Social recommendation addresses data sparsity and cold-start problems by incorporating users’ social networks into collaborative filtering.
- Prior Works and Limitations: Previous methods often equate friendship with preference similarity, despite social networks containing close, casual, and event friends.
- Prior Works and Limitations: Prior models commonly represent friends’ influences with constant weights or fixed constraints, overlooking effects that vary with specific items and contexts.
- Motivations and Rationales: The paper distinguishes user-side homophily and influence from analogous item-to-item effects, yielding four social effects that can jointly affect a user’s decision.Homophily is associated with static preference or attractiveness, whereas influence is context-dependent and dynamic.
- Methodologies and Results: DANSER uses dual GATs to model static and dynamic effects in user and item domains, with item information extending the social-effect representation.
- Methodologies and Results: A contextual multi-armed-bandit policy dynamically weights the four cross-domain interactions for each targeted user-item pair.The fusion strategy optimizes a reward related to predicted loss.
2 PRELIMINARY AND BACKGROUND
The paper formulates recommendation from user-item interactions and a trust or friendship graph, covering both implicit-click and explicit-rating settings.
- Problem Formulation: The interaction matrix R records clicks for implicit feedback and rating values for explicit feedback, with zero indicating an unobserved interaction.
- Raw Input and Item Implicit Network: The user social network is represented as a graph whose edges connect users, while interaction frequencies serve as edge features reflecting link strengths.
- Problem Formulation: The social recommendation task estimates unobserved interactions, including click probabilities for implicit feedback or ratings for explicit feedback.
3 METHODOLOGIES
DANSER models multifaceted social effects across user and item graphs using dual attention networks, then fuses their interactions with a context-dependent policy. Its framework combines static and dynamic factors to produce recommendation scores.
- Raw Input and Item Implicit Network: DANSER builds an item implicit network by linking items whose shared-user relevance exceeds a fixed threshold τ.The network complements the user social network and is derived from users who clicked or rated both items.
- Embedding Layer: Users and items receive low-dimensional embeddings, while interaction frequencies between users serve as edge features.User-specific and item-specific embeddings capture inherent preference or attribute factors; history-based embeddings additionally represent cross-domain information.
- Dual GCN/GAT Layer: Dual GATs model static homophily effects and context-aware influence effects for both users and items.The user branch produces static and dynamic preference factors, while the item branch produces analogous static and dynamic attribute factors.
- Policy-Based Fusion Layer: A policy-based fusion layer dynamically weights the four interacted features for each user–item pair.The policy uses the user and candidate-item context to allocate weights, reflecting that homophily and influence can matter differently across pairs.
- Discussions: DANSER’s global and local views select different influential neighbors when candidate-item similarity changes.In the illustrated cases, attention shifts toward u2 for one similar item and u3 for another, supporting context-sensitive neighbor selection.
4 EXPERIMENTS
The experiments evaluate DANSER against state-of-the-art recommendation and social recommendation models, test the necessity of its key components, examine hyper-parameter effects, and assess interpretation of four social effects.
- The experiments address four questions covering comparative performance, component necessity, hyper-parameter sensitivity, and interpretation of four social effects.
4.1 Experiment Setup
The evaluation uses Epinions and WeChat Top Story with different feedback types, data splits, metrics, comparison models, ablations, and implementation settings.
- Data sets: DANSER is evaluated on Epinions, a benchmark with explicit ratings and directed trust relationships, and WeChat Top Story, a commercial article-clicking dataset with implicit feedback.Epinions uses ratings from 1 to 5, whereas WeChat records whether users clicked articles and has bidirectional friendships.
- Data sets: Epinions uses 80% of user-item interactions for training and 20% for testing, while WeChat trains on the first 90% of chronologically ordered clicks.For WeChat, negative samples are uniformly sampled to balance positive and negative examples per user.
- Implementation: The implementation uses TensorFlow on a GTX 1080 GPU, with dataset-specific hyper-parameters including embedding dimension D = 10 and batch size B = 64 for Epinions.Additional settings include dropout ratio τ = 0.5, regularization coefficient λ = 0.001, and sample size F = 30.
- Competitors: The comparison includes SVD++, DELF, TrustPro, TrustMF, TrustSVD, NSCR, and SREPS, spanning collaborative-filtering and social-recommendation approaches.SVD++ and DELF use user-item interactions, while the social recommendation models also use user friendships.
- Ablation study: Ablations remove or replace dual GATs and compare policy-based fusion with max- and average-pooling variants.DualEMB removes neighboring-node convolution, DualGCN replaces GAT with GCN, and userGAT or itemGAT removes one domain’s GATs.
- Evaluation Protocol: Epinions is evaluated with MAE and RMSE, while WeChat uses Precision@k and AUC; each experiment is repeated ten times and averaged.Lower MAE and RMSE are better, whereas higher P@10 and AUC are better.
4.2 Comparative Results: RQ1
DANSER outperforms the comparative recommendation and social recommendation methods on both Epinions and WeChat, with reported improvements in MAE and AUC. The comparisons also indicate that model choice and social-information modeling affect performance across feedback settings.
- 2.87% MAE improvement on Epinions and 4.48% AUC improvement on WeChat are reported over the best competitors.The paper states that DANSER outperforms the comparative methods and achieves these improvements for social recommendation.
- TrustMF and TrustSVD perform well on Epinions with explicit feedback but poorly on WeChat with implicit feedback.
- NSCR performs better than other linear methods, indicating the relevance of neural networks for capturing complex non-linear information.
- DELF, despite using no social information, outperforms some social methods including TrustSVD and NSCR.The paper attributes this possibility to DELF’s dual embedding and neural interaction layer, alongside limited use of social information in some competing methods.
4.3 Ablation Study: RQ2
The ablation study finds that dual GATs and policy-based fusion each improve recommendation performance, while GAT-based attention is especially useful for users with many friends.
- Policy-based fusion and policy-gradient training substantially improve accuracy, especially on Epinions.The policy unit also outperforms straightforward concatenation in the DualEMB comparison.
- As clicked items increase, DANSER’s MAE/AUC variation narrows; as friend counts increase, DANSER improves while DualGCN declines.The authors attribute this pattern possibly to GAT filtering misleading neighbors more attentively than GCN.
4.4 Parameter Sensitivity: RQ3
Parameter sensitivity analysis shows that regularization, dropout, embedding dimension, and sample size affect DANSER’s MAE, with balanced settings needed for accuracy and complexity.
- Regularization, dropout, embedding dimension, and sample size all influence DANSER’s MAE on Epinions.The study reports that other hyper-parameters have little impact and are omitted.
- Excessive regularization prevents the model from focusing on minimizing recommendation loss.
- Too much dropout hinders training, whereas too little dropout can reduce its intended regularization effect.
- An appropriate sample size balances model complexity and accuracy.
4.5 Case Study: RQ4
The case study illustrates how DANSER’s four GATs and policy weights expose both context-dependent social effects and the dominant combinations influencing each user-item pair.
- Homophily GAT weights remain unchanged for the same user or item, whereas influence GAT weights vary across user-item pairs.
- Influence GATs assign larger weights to nodes sharing the same labels under a specific context.
- Policy weights select the most dominant interacted features from the four GAT representations.
- Interpretability operates at two levels: policy weights identify dominant domain-level effect combinations, while GATs identify influential friends and related items.
5 RELATED WORKS
Related work spans social recommendation, graph convolution and attention networks, and dual-model designs, while DANSER combines these strands to model multifaceted effects.
- Social Recommendation: Social recommendation methods use trust propagation, regularization, matrix factorization, deep neural networks, and network embedding to exploit social information.
- Social Recommendation: Earlier approaches often capture linear graph information, motivating methods that learn more complex graph representations.
- Graph Convolution/Attention Network: GCN and GAT encode complex graphs into low-dimensional representations, with GAT extending GCN through attention and providing better performance according to prior work.
- Dual Mechanism: Prior dual structures combine dual tasks, global and local graph consistencies, or user and item embeddings; DANSER differs through dual GAT-based multifaceted social-effect modeling.
6 CONCLUSION
DANSER uses dual graph attention networks to learn multifaceted social-effect representations across friend users and related items, improving recommendation accuracy. The work also points to richer item networks and extensions beyond recommendation.
- DANSER uses two dual graph attention networks to learn deep representations of social effects in recommender systems.
- The model collaboratively captures homophily and influence effects from both friend users and related items.
- Comparative experiments and ablation studies on a benchmark dataset and a commercial dataset showed significantly improved recommendation accuracy.
- Future extensions include enriching item implicit networks with attribute features or knowledge-graph connections when available.
- The framework could also extend to information-retrieval tasks such as question answering by modeling effects from related questions and similar answers.