Source-linked AI summary
Session-based Social Recommendation via Dynamic Graph Attention Networks
Weiping Song, Zhiping Xiao, Yifan Wang, Laurent Charlin, Ming Zhang, Jian Tang
TL;DR
Online-community recommendation must handle dynamic user interests and context-dependent influence from friends. The paper proposes a dynamic-graph-attention recommender combining recurrent session modeling with graph attention, and reports superior performance on three real-world data sets. Its evaluation also leaves session-segmentation optimization for future work.
Problem
Recommendation in online communities must address changing user interests and social influences whose relevant friends vary by context.
Method
The model uses an RNN to represent current session behavior and graph attention to combine user and friend representations dynamically.
Results
The model outperforms well-known competitive baselines on Douban, Delicious, and Yelp data sets.
Takeaways & Limitations
Dynamic user behavior and dynamic social influence can be modeled together for session-based social recommendation.
Takeaways & Limitations
The paper leaves systematic optimization of session segmentation for future work.
Abstract
from arXiv · showhide
Online communities such as Facebook and Twitter are enormously popular and have become an essential part of the daily life of many of their users. Through these platforms, users can discover and create information that others will then consume. In that context, recommending relevant information to users becomes critical for viability. However, recommendation in online communities is a challenging problem: 1) users' interests are dynamic, and 2) users are influenced by their friends. Moreover, the influencers may be context-dependent. That is, different friends may be relied upon for different topics. Modeling both signals is therefore essential for recommendations. We propose a recommender system for online communities based on a dynamic-graph-attention neural network. We model dynamic user behaviors with a recurrent neural network, and context-dependent social influence with a graph-attention neural network, which dynamically infers the influencers based on users' current interests. The whole model can be efficiently fit on large-scale data. Experimental results on several real-world data sets demonstrate the effectiveness of our proposed approach over several competitive baselines including state-of-the-art models.
1 INTRODUCTION
Online-community recommendation must account for changing user interests and social influences that vary with context. The paper proposes a dynamic-graph-attention approach and reports effectiveness over competitive baselines on real-world data.
- Online communities require recommendation systems that surface relevant information because users create, share, and consume content on these platforms.
- User interests change over time, while friends can influence recommendations through short- and long-term preferences.
- Different friends may influence a user in different sessions because social influence is context-dependent.
- Prior recommendation work modeled dynamic interests or social influence separately, but did not combine both signals.
- The proposed model uses an RNN for session behavior and graph attention to infer influential friends from current interests.
- Experimental results on several real-world data sets show the model outperforms well-known competitive and state-of-the-art baselines.
2 RELATED WORK
Related work covers dynamic-interest recommendation, social recommendation, and graph convolutional networks. The paper distinguishes its approach by modeling both evolving node features and context-dependent attention in a dynamic graph.
- Relevant research includes dynamic user-behavior recommenders, social recommenders, and convolutional networks for graph-structured data.
- Earlier temporal recommendation models commonly use factorization and assume that user interests change slowly and smoothly.
- Social recommendation methods model friend influence with latent-factor regularization, trust weighting, transfer learning, or related mechanisms.
- These social models generally assume influence is uniform across friends rather than context-dependent.
- Graph convolutional networks extend convolutional modeling from two-dimensional grids to more general graph structures.
- The proposed dynamic-graph-attention network models evolving node features and attention relationships as context changes over time.
3 PROBLEM DEFINITION
Session-based social recommendation predicts a user’s next items from current-session behavior and social information. The formalization captures dynamic interests alongside context-dependent influences from friends.
- Classical recommendation often ignores consumption order, whereas session-based recommendation models sequences of user behaviors.
- The session-based recommendation task recommends items likely to interest a user at the next time step from a new session sequence.
- In online communities, a user’s current interests are influenced by friends as well as historical behavior.
- Social influence is context-dependent: users may rely on different friends for laptops, cameras, or other situations.
- Session-based social recommendation uses both a user’s dynamic interests and the social influences of her friends.
4 DYNAMIC SOCIAL RECOMMENDER SYSTEMS
DGRec combines dynamic session interests with friends’ short- and long-term preferences, using graph attention to infer context-dependent social influence for recommendations.
- 4.1–4.2 Dynamic Interests: DGRec models a target user’s current session with an RNN and combines it with friends’ short- and long-term preferences.Friends’ recent sessions are encoded with an RNN, while long-term preferences use individual embeddings.
- 4.3 Context-dependent Social Influences: The graph-attention network represents users and friends as graph nodes and propagates their dynamic features through friendship edges.The target user’s and friends’ representations provide node features for message passing.
- 4.3 Context-dependent Social Influences: Attention weights determine each friend’s influence on the target user, conditioned on the current context, rather than treating neighbors equally.The model contrasts attention-based propagation with static graph-convolution strategies that assign equal treatment to neighbors.
- 4.3–4.4 Recommendation: After L attention layers, the target node yields a social-influenced representation that is combined with the user’s current interest.A fully connected layer produces the final user representation before recommendation scoring.
- 4.4–4.5 Training: The model predicts the next item with a softmax and trains by maximizing the log-likelihood of observed items using gradient descent.The item embedding and total item set are used in the softmax probability.
- 4.3.2 Graph-Attention Network: Mean aggregation outperforms element-wise max-pooling, but both context-independent alternatives are inferior to the proposed attention model.This comparison supports using attention rather than fixed, context-independent friend aggregation.
5 EXPERIMENTS
The experiments evaluate DGRec on real-world data sets and report strong performance, component usefulness, and context-dependent friend weighting.
- DGRec significantly outperforms all seven compared methods under all experimental settings.
- Ablation studies demonstrate the usefulness of DGRec’s different components.
- Fitted-model analysis shows that attention contextually weighs friends’ influences.
5.1 Experimental Setup
The study evaluates recommenders on three online-community data sets using ranking metrics, standardized procedures, and comparisons spanning classical, social, and session-based methods.
- Data Sets: The evaluation uses Douban, Delicious, and Yelp data collected from three online communities.Douban and Yelp use review consumption observations, while Delicious evaluates personalized tag recommendations for bookmarks.
- Data Sets: Other session lengths were tested, but a systematic study of session-segmentation optimization is left for future work.Preliminary results showed consistent outperformance over Session-RNN and NARM for other session lengths.
- Evaluation Protocol: Testing reserves the last d days and splits held-out sessions equally into validation and test sets, with d = 180, 50, and 25 for Douban, Yelp, and Delicious.
- Competing Models: DGRec is compared with classical, social, and session-based recommendation methods.The compared classes respectively omit social and temporal factors, model context-independent social influence, or model user interests within sessions.
- Evaluation Metrics: Models are evaluated using Recall@K and NDCG, with K = 20 for Recall@K.NDCG is averaged over testing examples and uses the rank of a positive item.
- Implementation Details: Implementation uses TensorFlow, mini-batch training, Adam optimization, 100-dimensional representations, and empirically selected recurrent and neighborhood settings.Neighbor sampling is used to reduce graph-training difficulty and cost; friends’ short-term interests use their most recent sessions.
- Quantitative Results: Table 2 reports quantitative results for DGRec and the compared baselines across three data sets and both metrics.
5.2 Quantitative Results
DGRec achieves the best performance among the evaluated algorithms by combining dynamic user interests with context-dependent social influences.
- DGRec achieves the best performance among all algorithms by modeling dynamic user interests and context-dependent social influences.Its improvement over RNN-Session and NARM is more significant than SoReg’s improvement over BPR-MF.
- RNN-Session significantly outperforms ItemKNN and BPR-MF, while NARM becomes the strongest baseline by modeling the user’s main purpose.
- Social recommenders improve over BPR-MF in most cases, but their improvement is marginal when social influence is context-independent.
- On Douban, BPR-MF performs poorly because matrix-factorization methods tend to recommend items users previously consumed.Douban users typically consume each item only once, unlike users on Delicious and Yelp.
5.3 Variations of DGRec
DGRec benefits from combining individual session interests with context-dependent social influence, using both short- and long-term friend preferences and two graph-convolution layers.
- Self v.s. Social: DGRec combines the target user’s current session with context-dependent social influence, whereas either source alone significantly reduces performance.The individual-only variant is DGRecself, and the social-only variant is DGRecsocial.
- Self v.s. Social: DGRecself consistently outperforms DGRecsocial across Douban, Delicious, and Yelp, indicating greater overall impact from individual interests.
- Short-term v.s. Long-term: On Douban, friends’ short-term interests drastically outperform long-term interests and perform comparably to the full model.
- Short-term v.s. Long-term: On Delicious, long-term friend interests yield more accurate predictions than short-term interests.The paper attributes this pattern to users’ relatively static interests on Delicious.
- Convolution Layers: A single graph-convolution layer causes a significant performance decline, showing that friends’ friends are important for recommendations.
- Convolution Layers: Three graph-convolution layers provide little benefit and reduce performance on Douban and Delicious, so two layers are sufficient for these data sets.The paper notes a small Yelp improvement and possible overfitting or noise from higher-order friends.
5.4 Exploring Attention
DGRec’s attention mechanism identifies which friends matter across contexts and sessions. Attention varies more across sessions than within a session, supporting context-dependent social influence.
- DGRec weighs friends differently according to the target user’s current session and interests.The model uses attention to select context-dependent influencers rather than treating all social links equally.
- Figure 5 maps attention weights across eight sessions and across the item sequence in session #7, with friends on the y-axis.The left plot shows inter-session attention, while the right plot shows intra-session attention.
- Figure 6 reports empirical inter-session and intra-session attention-variance distributions using 20 discretized intervals.The distributions are computed over attention inferred on the Douban test set.
- The target user allocates attention to different friends across different sessions, indicating that social influence depends on current context.Friend #8 receives little attention in all sessions, showing that social links do not necessarily produce influence.
- Intra-session attention variance is lower on average than inter-session variance, consistent with the model’s assumption about session-specific friend reliance.The macro-level analysis compares attention variation across all users.
6 CONCLUSIONS
The paper presents a graph-convolutional model for session-based social recommendation. It combines current-interest representations with attention-weighted friend representations and reports superior performance across three real-world datasets.
- The model learns individual user representations from current interests and aggregates friends’ representations with an attention-based graph convolutional network.The combined representation and the user’s original representation are used to form item recommendations.
- Experimental results on three real-world datasets demonstrate superiority over several state-of-the-art models.
- Future work will explore user and item features indicative of preferences and further improve recommender-system performance for online communities.