Source-linked AI summary
Social Recommendation with Self-Supervised Metagraph Informax Network
Xiaoling Long, Chao Huang, Yong Xu, Huance Xu, Peng Dai, Lianghao Xia, Liefeng Bo
TL;DR
Social recommendation uses social information to address collaborative filtering’s data sparsity, but existing systems lack explicit item-side semantic relations and face challenges modeling heterogeneous dependencies. SMIN combines metapath-guided heterogeneous graph encoding with self-supervised graph mutual-information learning for user-item interactions. Experiments report that it outperforms competing state-of-the-art alternatives across three datasets.
Problem
Social recommendation seeks to alleviate collaborative filtering’s data sparsity, but existing systems lack explicit item-semantic encoding and face challenges modeling heterogeneous, high-order user- and item-side relations.
Source: §1 · pp. 1–2Method
SMIN uses metapath-guided heterogeneous graph encoding and cross-metapath attention, then applies self-supervised graph mutual-information learning to incorporate relational structure into user-item interaction modeling.
Source: §1 · p. 2, §3.3.3 · p. 5Results
SMIN consistently outperforms other alternatives on both metrics across three datasets.
Source: §4.2 · p. 7Takeaways & Limitations
The authors conclude that SMIN offers an advantage for knowledge-aware social recommendation by learning user- and item-side social and knowledge-graph relationships.
Source: §6 · p. 9
Abstract
from arXiv · showhide
In recent years, researchers attempt to utilize online social information to alleviate data sparsity for collaborative filtering, based on the rationale that social networks offers the insights to understand the behavioral patterns. However, due to the overlook of inter-dependent knowledge across items (e.g., categories of products), existing social recommender systems are insufficient to distill the heterogeneous collaborative signals from both user and item sides. In this work, we propose a Self-Supervised Metagraph Infor-max Network (SMIN) which investigates the potential of jointly incorporating social- and knowledge-aware relational structures into the user preference representation for recommendation. To model relation heterogeneity, we design a metapath-guided heterogeneous graph neural network to aggregate feature embeddings from different types of meta-relations across users and items, em-powering SMIN to maintain dedicated representations for multi-faceted user- and item-wise dependencies. Additionally, to inject high-order collaborative signals, we generalize the mutual information learning paradigm under the self-supervised graph-based collaborative filtering. This endows the expressive modeling of user-item interactive patterns, by exploring global-level collaborative relations and underlying isomorphic transformation property of graph topology. Experimental results on several real-world datasets demonstrate the effectiveness of our SMIN model over various state-of-the-art recommendation methods. We release our source code at https://github.com/SocialRecsys/SMIN.
1 INTRODUCTION
SMIN addresses gaps in social recommendation by jointly modeling heterogeneous social and item-knowledge relations, then using self-supervised graph learning to incorporate collaborative signals into user-item interaction representations.
1 INTRODUCTION
- Social information has been incorporated into collaborative filtering to address data sparsity, with earlier approaches using regularization or trust-based latent representations and later work using attention and graph propagation.Examples include SoReg, TrustMF, SAMN, EATNN, GraphRec, DANSER, and DiffNet.Source: §1 · p. 1
- Existing social recommenders lack explicit encoding of semantic relatedness between items, while the joint user- and item-side signals pose heterogeneity and high-order modeling challenges.Item categories, functionality, and review connections are examples of item-side knowledge; the framework must also preserve local and global graph structure.Source: §1 · pp. 1–2
- SMIN uses metapath-guided message passing to preserve relation-specific user and item representations, then attention-based cross-metapath aggregation to summarize their semantics.The metapaths capture social homophily, item knowledge dependencies, and behavioral co-interaction patterns.Source: §1 · p. 2
- A self-supervised graph mutual-information architecture transfers user-user and item-item relations into interaction modeling and targets local and global collaborative signals.The proposed augmentation considers global substructure context, node features, and transformation characteristics of user-item relations.Source: §1 · p. 2
- The paper emphasizes jointly modeling social- and knowledge-aware relations from both user and item domains under self-supervised learning.Its stated aim is to learn user preferences through local and global collaborative similarities.Source: §1 · p. 2
- The paper reports that experiments on three real-world datasets show SMIN outperforming state-of-the-art recommendation techniques.The introduction presents this as evidence for the framework’s effectiveness.Source: §1 · p. 2
2 PRELIMINARY
The preliminary section defines the interaction and side-information graphs used to formulate knowledge-aware social recommendation as prediction of unobserved user-item behavior.
2 PRELIMINARY
- The interaction matrix records whether each user-item pair is observed, and the user social graph connects socially related users.The section introduces users, items, and binary interaction indicators before defining the social graph.Source: §2 · p. 2
- The item relational graph represents item dependencies as relation triples, including category membership and co-interaction.Its triples connect a head item and tail item through a relation type.Source: §2 · p. 3
- The collaborative heterogeneous graph unifies user-item interactions, user social links, and item relations; the task is to predict unknown user-item behavior.The unified graph uses node and edge types to represent heterogeneous structure.Source: §2 · p. 3
3 METHODOLOGY
SMIN combines metapath-guided heterogeneous graph encoding, cross-metapath aggregation, and self-supervised graph mutual-information learning to model social and knowledge-aware relations in recommendation. Its learning process integrates these representations with user-item interaction modeling.
3.1 Meta-Relation Heterogeneity Encoding
- Metapaths encode social and behavioral user relations, knowledge-aware user dependencies, and behavior- and knowledge-based item dependencies.The user relations include established social connections and behavioral influence; item relations include interaction-based attractiveness and knowledge-graph connections.Source: §3.1.1 · p. 3
- Metapath-guided graph layers aggregate features from metapath-based neighbors to generate user and item embeddings.Embeddings are updated over multiple graph layers, then cross-layer representations are concatenated.Source: §3.1.2 · p. 4
3.2 Cross-metapath Aggregation Network
- An attention-based aggregation layer learns normalized importance weights for metapaths and combines their heterogeneous user- and item-side semantics.The same aggregation mechanism is applied to user- and item-wise meta-relations.Source: §3.2 · p. 4
3.3 Self-Supervised Mutual Information Learning
- Substructure-aware global context uses an adjacency matrix built from different-order relations to inject graph-level context into latent representations.The k-order matrix incorporates k-hop neighbors; the described examples construct second- and third-order matrices from lower-order relations.Source: §3.3.2 · p. 5
- Global-context mutual information learning pairs local node embeddings with high-level graph representations to capture user-item patterns from local to global.The framework maximizes mutual information between node embeddings and corresponding graph representations.Source: §3.3 · p. 4, §3.3.2 · p. 5
- Topology-aware mutual information learning models node and edge transformations to preserve structural characteristics and user-item interactions.The edge transformation uses reconstructed user-item connections generated from the encoded user and item representations.Source: §3.3.3 · p. 5
3.4 The Learning Process of SMIN
- SMIN jointly optimizes heterogeneous relationship learning and graph-based interaction modeling using a Bayesian personalized ranking objective.The objective balances the mutual-information losses from different views with regularization.Source: §3.4.1 · p. 5
- SMIN's overall computational complexity is O((|E| + |E_u-v|) × L × d), which the authors describe as comparable to most existing social recommendation approaches.The analysis identifies cross-layer propagation and the final prediction layer as component costs.Source: §3.4.2 · p. 6
4 EVALUATION
The evaluation tests SMIN on three real-world datasets against recommendation baselines and examines its components, meta-relations, hyperparameters, and learned embeddings. SMIN outperforms the alternatives in the reported comparisons, while the analyses describe how its design choices relate to performance.
4.1 Experimental Settings
- Experiments use Ciao, Epinions, and Yelp, with leave-one-out evaluation, one positive paired with 99 negatives, HR and NDCG metrics, and a range of social, graph, and knowledge-aware baselines.The item knowledge relations are generated from item categories.Source: §4.1.1 · p. 6, §4.1.2 · p. 6, §4.1.3 · p. 6
4.2 Performance Comparison (RQ1)
- Across three datasets, SMIN consistently outperforms competing methods on HR@10 and NDCG@10, and also shows superiority across different Top-N settings on Epinions.Table 2 reports the three-dataset comparison; Table 3 reports ranking performance as Top-N varies.Source: Table 2 · p. 7, §4.2 · p. 7, Table 3 · p. 7
4.3 Ablation Study of SMIN Framework (RQ2)
- Ablations report improved performance from heterogeneous relation modeling, multi-view mutual information, global context injection, topology-aware relational learning, and differentiated metapath weighting.The authors relate the benefit of metapath weighting to user-user and item-item relations contributing differently to user-item interaction modeling.Source: §4.3 · pp. 7–8
4.4 Effect Analyses of Meta-Relations (RQ3)
- The full set of meta-relations achieves the best recommendation accuracy; removing social, behavior-level, item-semantic, or knowledge-aware relations reduces the reported performance.The analyses associate these relations with social influence, user-item dependencies, item semantic relatedness, and item-specific preference.Source: Figure 3 · p. 8, §4.4 · p. 8
4.5 Hyperparameter Study (RQ4)
- Performance saturates around d=64, is best at L=2, and can worsen with larger embeddings, L=3, or higher-order k=3 context.The authors attribute the degradation at larger d to overfitting and note that k=3 may introduce noise and irrelevant dependencies; Figure 4 plots changes relative to the first data point.Source: §4.5 · p. 8, Figure 4 · p. 8
4.6 User/Item Embedding Visualizations (RQ5)
- The t-SNE visualization shows better clustering for SMIN embeddings than for GraphRec and HERec, as well as clearer separation of users interacting with different items.The authors interpret these observations as evidence that the learned embedding space preserves user-item relationships and bridges user- and item-domain semantics.Source: §4.6 · pp. 8–9, Figure 5 · p. 9
5 RELATED WORK
Prior work models social influence and user-item graph structure, but often omits item-side knowledge or does not combine it with multifaceted social signals. SMIN connects these strands through metapath-guided heterogeneous graph learning and self-supervised relational knowledge.
5.1 Social Recommendation Techniques
- Social recommendation uses attention and graph neural networks to model user influence, but many methods omit item-domain knowledge from user-item interaction encoding.Examples include DiffNet, DANSER, and GraphRec.Source: §5.1 · p. 9
5.2 GNNs for Recommendation
- Graph-based recommendation progresses from interaction-graph aggregation in PinSage and NGCF to LightGCN’s streamlined neighborhood aggregation; this work combines metapath-guided graph learning with mutual-information modeling.The resulting framework integrates a metapath-guided graph neural module and substructure-aware mutual-information modeling for social recommendation.Source: §5.2 · p. 9
5.3 Self-Supervised Learning
- Unlike prior graph self-supervision based on reconstruction or mutual-information exploration, this work uses user- and item-side relational knowledge as self-supervised information for social recommendation.The passage states this can enhance user/item representations with knowledge-aware global information.Source: §5.3 · p. 9
5.4 Knowledge-aware Recommender Systems
- Knowledge-aware recommenders transfer item-side structural knowledge into interaction modeling, while SMIN also explores multifaceted social signals; examples of prior methods include KGAT and KGNN-LS.The cited methods investigate structural knowledge through embedding propagation.Source: §5.4 · p. 9
6 CONCLUSION
SMIN learns social- and knowledge-dependent user and item relationships with a metapath-guided heterogeneous graph neural network, augmented by self-supervised learning of graph structure. Experiments on several real-life datasets show an advantage over various state-of-the-art recommendation models.
6 CONCLUSION
- SMIN learns social- and knowledge-dependent user and item relationships through a metapath-guided heterogeneous graph neural network.The model is proposed for knowledge-aware social recommendation.Source: §6 · p. 9
- A self-supervised learning framework augments graph-structural modeling to preserve complex relationships among users and items.Source: §6 · p. 9
- Experiments on several real-life datasets demonstrate SMIN’s advantage over various state-of-the-art recommendation models.Source: §6 · p. 9