Source-linked AI summary
Knowledge-aware Coupled Graph Neural Network for Social Recommendation
Chao Huang, Huance Xu, Yong Xu, Peng Dai, Lianghao Xia, Mengyin Lu, Liefeng Bo, Hao Xing, Xiaoping Lai, Yanfang Ye
TL;DR
Social recommendation must account for item dependencies, heterogeneous interaction types, and temporal behavior beyond user social connections alone. KCGN addresses these gaps with a coupled, mutual-information-based graph framework and relation-aware temporal encoding, achieving substantial gains over state-of-the-art baselines across real-world settings.
Problem
Social recommendation methods often overlook item dependencies, heterogeneous interaction types, and dynamic user-item behavior, limiting the modeled information about user preferences.
Method
KCGN jointly learns user-user and item-item structures with a mutual information-based coupled graph neural network and relation-aware temporal interaction encoding.
Results
KCGN achieves substantial gains over state-of-the-art baselines and consistently obtains the best HR@10 and NDCG@10 performance across recommendation scenarios.
Takeaways & Limitations
KCGN provides a unified framework for knowledge-aware, multi-typed, and temporally informed social recommendation.
Abstract
from arXiv · showhide
Social recommendation task aims to predict users' preferences over items with the incorporation of social connections among users, so as to alleviate the sparse issue of collaborative filtering. While many recent efforts show the effectiveness of neural network-based social recommender systems, several important challenges have not been well addressed yet: (i) The majority of models only consider users' social connections, while ignoring the inter-dependent knowledge across items; (ii) Most of existing solutions are designed for singular type of user-item interactions, making them infeasible to capture the interaction heterogeneity; (iii) The dynamic nature of user-item interactions has been less explored in many social-aware recommendation techniques. To tackle the above challenges, this work proposes a Knowledge-aware Coupled Graph Neural Network (KCGN) that jointly injects the inter-dependent knowledge across items and users into the recommendation framework. KCGN enables the high-order user- and item-wise relation encoding by exploiting the mutual information for global graph structure awareness. Additionally, we further augment KCGN with the capability of capturing dynamic multi-typed user-item interactive patterns. Experimental studies on real-world datasets show the effectiveness of our method against many strong baselines in a variety of settings. Source codes are available at: https://github.com/xhcdream/KCGN.
Introduction
The paper identifies three gaps in social recommendation: missing item dependencies, limited modeling of heterogeneous interactions, and insufficient attention to temporal dynamics. KCGN addresses these gaps by jointly modeling user-user and item-item relations, relation-specific interactions, and temporal information.
- Research gaps: Existing social recommender systems primarily model user-user and user-item relations while overlooking item-wise semantic dependencies.Such dependencies include category, functionality, or complementary relations that can enrich item knowledge.
- Research gaps: Many methods assume a single interaction type, although behaviors such as page views, favorites, and purchases have distinct and inter-dependent meanings.Encoding these heterogeneous patterns is important for representing users’ complex interests.
- Research gaps: Most recent approaches ignore the dynamic nature of user-item interactions, while sequential methods remain limited by singular interaction modeling.The paper motivates incorporating time alongside multiple interaction types.
- Technical challenges: Graph neural aggregation can capture local user proximity but is insufficient for broader graph-context and high-order relational awareness.The paper identifies global graph-structure awareness as a technical challenge.
- Proposed approach: KCGN jointly captures user-user and item-item relations through a coupled graph neural network, preserving knowledge-aware cross-item relations in user embeddings.This joint modeling is intended to enhance social-aware preference representations.
- Proposed approach: A relation-aware graph neural module encodes multi-typed interactions and incorporates temporal information into message passing.The model is evaluated on three real-world datasets, including scalability and cold-start settings.
Problem Definition
The paper formalizes social recommendation using multi-typed user-item interactions, timestamps, a user social graph, and an item inter-dependency graph. The task is to learn a predictive function for future user-item interactions from these inputs.
- Interaction representation: The problem setting contains I users and J items, with multiple interaction types represented between users and items.The interaction types include examples such as page view, purchase, like, and dislike.
- Interaction representation: The multi-typed interaction tensor X ∈ R^I×J×K records whether each user-item pair exhibits interaction type k.Its entries are binary indicators for the presence or absence of a typed interaction.
- Temporal representation: The temporal tensor T has the same dimensions as X and records timestamps for corresponding interactions.This representation supplies temporal information for modeling interaction dynamics.
- Graph structures: The user social graph G_u represents users as nodes and social connections as edges.An edge connects two users when they are socially connected.
- Graph structures: The item inter-dependency graph G_v represents relations such as shared categories, similar functionality, or common typed interactions by users.These edges encode item-wise dependencies used as relational knowledge.
- Task formulation: Given X, G_u, and G_v, the task is to forecast future user-item interactions with a predictive function.Figure 1 presents the architecture for modeling multi-typed interactive patterns, including element-wise addition.
Methodology
KCGN combines relation-aware message passing for heterogeneous, time-stamped user-item interactions with coupled learning over user social and item inter-dependency graphs. Mutual-information objectives preserve local node characteristics alongside global graph dependencies, while gating summarizes type-specific item representations.
- Multi-typed Interactive Pattern Modeling: KCGN constructs a multi-typed relation graph with type-specific item subvertices to encode heterogeneous user-item interactions.The graph contains I + J · K vertices, comprising users and type-aware item vertices.
- Multi-typed Interactive Pattern Modeling: Message construction combines user and type-specific item embeddings, an encoding function, and degree-based decay to normalize propagated influence.The decay factor uses neighboring-node counts for users and type-specific item connections.
- Multi-typed Interactive Pattern Modeling: Temporal context encoding maps interaction timestamps into time slots and adds sinusoidal relative-time embeddings to interaction edges.These temporal embeddings are incorporated into message propagation to model dynamic dependencies across interaction types.
- Multi-typed Interactive Pattern Modeling: High-order aggregation propagates messages across graph layers, then uses gating to differentiate the importance of type-specific interaction patterns when summarizing item embeddings.The resulting representations concatenate layer-wise user and item embeddings before type-aware item summarization.
- Knowledge-aware Coupled Graph Neural Module: The coupled graph module jointly models local user-social and item-interdependency structures with global graph-substructure awareness through mutual-information learning.It mean-pools node embeddings into graph-level representations, contrasts matched and shuffled node-graph pairs, and maximizes mutual information through a contrastive loss.
- Model Optimization: KCGN optimizes a combined objective containing pairwise BPR loss, multi-typed interaction encoding, knowledge-aware relation learning, mutual-information maximization, and L2 regularization.Its message passing costs O(|E| × d), while transformations cost O((I + J · K) × d^2).
Evaluation
KCGN is evaluated across accuracy, interaction types, sparsity, representation quality, sensitivity, and efficiency. It consistently performs strongly, with joint relation modeling particularly beneficial for sparse users and multi-typed interactions.
- Overall accuracy: KCGN consistently achieves the best HR@10 and NDCG@10 across recommendation scenarios, integrating user-user, item-item, and multi-typed user-item relations.
- Target interactions: KCGN also improves prediction of target like or purchase interactions by explicitly modeling different user-item interaction signals.KGAT does not explicitly differentiate type-specific interaction patterns.
- Ablation analysis: The joint KCGN model outperforms variants that remove multi-typed interactions, social relations, item dependencies, temporal context, or coupled relation encoders.KCGN-UI performs worse than KCGN-U and KCGN-I, supporting the relation aggregation functions.
- Sparsity: KCGN outperforms representative baselines in most sparsity groups, especially the sparsest Yelp and E-Commerce users.The authors attribute this to recursive aggregation of user and item side knowledge.
- Representation quality: KCGN embeddings preserve user-item relationships through clustering and provide better separation among users and their interacted items than NGCF+S.
- Sensitivity analysis: Two graph propagation layers outperform zero- and one-layer variants, whereas three layers introduce noise; increasing embedding size eventually causes overfitting.
- Efficiency: KCGN has competitive per-epoch running time compared with neural social recommendation baselines.
Related Work
Recent social recommendation work uses neural networks and graph-based approaches to combine social information with user-item interaction encoding.
- Attention mechanisms model influences between users in social recommendation systems.
- Recent efforts also explore graph neural networks for incorporating social relations into user-item interaction modeling.
Conclusion
KCGN integrates knowledge-aware item dependencies with social recommendation through coupled graph learning and mutual information, while modeling dynamic multi-typed interactions. Experiments on real-world datasets show substantial gains over state-of-the-art baselines.
- KCGN jointly learns user-user and item-item relation structures with a coupled graph neural network and mutual information-based neural estimator.
- A relation-aware graph encoder maintains dedicated representations for different interaction types while incorporating temporal information.
- KCGN achieves substantial gains over state-of-the-art baselines in extensive experiments on real-world datasets.