Source-linked AI summary
Hypergraph Contrastive Collaborative Filtering
Lianghao Xia, Chao Huang, Yong Xu, Jiashu Zhao, Dawei Yin, Jimmy Xiangji Huang
TL;DR
Graph-based collaborative filtering faces over-smoothing and scarce, noisy interaction supervision. HCCF combines local graph encoding, global hypergraph structure learning, and cross-view contrastive self-supervision. It reports consistent gains over state-of-the-art baselines, including a 24.7% average relative improvement over LightGCN in NDCG@20.
Problem
Graph-based CF is limited by over-smoothing that makes user vectors indistinguishable and by scarce, noisy interaction supervision.
Method
HCCF jointly models local collaborative relations and global hypergraph dependencies, using the two representations as complementary contrastive views.
Results
24.7% average relative improvement over LightGCN in NDCG@20 is reported, alongside gains over DHCF, MHCN, and SGL.
Takeaways & Limitations
Cross-view supervision provides self-data augmentation for sparse interactions while retaining local and global collaborative information.
Abstract
from arXiv · showhide
Collaborative Filtering (CF) has emerged as fundamental paradigms for parameterizing users and items into latent representation space, with their correlative patterns from interaction data. Among various CF techniques, the development of GNN-based recommender systems, e.g., PinSage and LightGCN, has offered the state-of-the-art performance. However, two key challenges have not been well explored in existing solutions: i) The over-smoothing effect with deeper graph-based CF architecture, may cause the indistinguishable user representations and degradation of recommendation results. ii) The supervision signals (i.e., user-item interactions) are usually scarce and skewed distributed in reality, which limits the representation power of CF paradigms. To tackle these challenges, we propose a new self-supervised recommendation framework Hypergraph Contrastive Collaborative Filtering (HCCF) to jointly capture local and global collaborative relations with a hypergraph-enhanced cross-view contrastive learning architecture. In particular, the designed hypergraph structure learning enhances the discrimination ability of GNN-based CF paradigm, so as to comprehensively capture the complex high-order dependencies among users. Additionally, our HCCF model effectively integrates the hypergraph structure encoding with self-supervised learning to reinforce the representation quality of recommender systems, based on the hypergraph-enhanced self-discrimination. Extensive experiments on three benchmark datasets demonstrate the superiority of our model over various state-of-the-art recommendation methods, and the robustness against sparse user interaction data. Our model implementation codes are available at https://github.com/akaxlh/HCCF.
1 INTRODUCTION
Graph-based collaborative filtering achieves strong recommendation performance but remains vulnerable to over-smoothing and scarce, noisy interaction supervision. HCCF addresses these issues by jointly learning local interaction and global hypergraph views with contrastive self-supervision.
- Motivation: Deeper graph propagation can make user vectors indistinguishable, mixing harmful noise into representations of individual preferences.Figure 1 illustrates this over-smoothing issue in graph-based CF models.
- Motivation: Sparse and skewed user-item interactions limit supervised CF representation learning, while observed connections can also be noisy or incomplete.The scarcity is especially problematic because many users interact with only a limited number of items.
- Approach: HCCF jointly learns parameterized hypergraph-guided user structures and the original interaction-graph encoder to improve user preference representations.The hypergraph component targets latent dependencies while the interaction graph preserves observed collaborative patterns.
- Approach: HCCF treats the explicit interaction graph and learned implicit hypergraph as contrastive views without random node or edge perturbations.This design avoids potentially dropping essential interactions used to encode users’ real interests.
- Results: 24.7% average relative improvement over LightGCN in NDCG@20 is reported, alongside gains over DHCF, MHCN, and SGL.The reported average relative improvements are 33.8% over DHCF, 20.2% over MHCN, and 14.8% over SGL.
2 PRELIMINARIES AND RELATED WORK
Prior CF methods use latent embeddings, graph propagation, hypergraphs, and contrastive learning to model user-item relations and high-order dependencies. HCCF combines these directions by modeling local collaborative similarities, global hypergraph dependencies, and complementary contrastive views.
- Collaborative Filtering: CF represents users and items as embeddings and estimates relevance to predict unobserved user-item interactions.The interaction matrix records observed consumption relationships between users and items.
- Collaborative Filtering: Matrix factorization, neural networks, and autoencoders provide alternative embedding or interaction-reconstruction functions for CF.NCF and DMF use multilayer perceptrons, while AutoRec and CDAE use autoencoder-based reconstruction.
- Graph-based CF: Graph-oriented approaches propagate information over multi-hop user-item topologies to capture high-order collaborative signals.NGCF and PinSage are cited as examples, while LightGCN modifies graph convolutional propagation.
- Hypergraph CF: Hypergraph recommender systems model complex high-order dependencies through node-hyperedge connections.HyRec, DHCF, and MHCN are cited as examples of hypergraph-based recommendation methods.
- Contrastive Learning: HCCF combines graph-based local similarity encoding, hypergraph-based global dependency learning, and complementary local-global contrastive views.The framework uses self-supervision to supplement collaborative-effect encoding under a hypergraph learning schema.
3 METHODOLOGY
HCCF combines local graph propagation with learned global hypergraph dependencies, then uses cross-view contrastive supervision to improve user and item representations. Residual aggregation and low-rank hypergraph parameterization address over-smoothing and computational cost while preserving complementary collaborative effects.
- Local Collaborative Relation Encoding: HCCF first encodes local collaborative similarities through graph-based message passing over the user-item interaction graph.Multiple propagation layers aggregate local neighborhood information for contextual user/item embeddings.
- Local Collaborative Relation Encoding: Residual connections preserve centric-node information during graph aggregation and alleviate GNN over-smoothing.The residual operation incorporates self-information between source and target nodes across graph layers.
- Hypergraph Global Dependency Learning: Hypergraph message passing uses hyperedges as intermediate hubs to capture global high-order relations among users and items without a hop-distance limit.The learned dependency matrices H (u) and H (v) provide the node-hyperedge structures used with graph-refined embeddings.
- Hypergraph Global Dependency Learning: Low-rank factorization parameterizes hypergraph structures as H (u) = E (u) · W (u) and H (v) = E (v) · W (v), adding O(H × d) time complexity.This reduces the parameter and computational burden of learning dense hypergraph adjacency matrices.
- Hypergraph Global Dependency Learning: Hierarchical hypergraph layers model higher-level hyperedge feature interactions using residual nonlinear transformations.The hyperedge embeddings are formed from learned node-hyperedge structures and user/item representations, then refined across c layers.
- Hypergraph-enhanced Contrastive Learning: Hypergraph-guided contrastive learning aligns local interaction-graph and global hypergraph views, providing auxiliary self-supervision for user and item representations.Edge dropout is applied to both structures, and the contrastive loss is integrated with CF loss in a unified objective.
4 EVALUATION
The evaluation compares HCCF with diverse recommender baselines on three datasets and examines its components, sparsity robustness, over-smoothing behavior, interpretation, and hyperparameters. HCCF consistently outperforms baselines, while analyses support contributions from hypergraph learning and cross-view contrastive supervision.
- Experimental Settings: Experiments use Yelp, MovieLens, and Amazon-book, with 7:1:2 train-validation-test splits, all-ranking evaluation, Recall@N, and NDCG@N.The comparison includes 15 baselines spanning matrix factorization, neural, autoencoder, GNN-based, and self-supervised recommendation methods.
- Overall Performance: HCCF consistently outperforms all baselines across datasets and evaluation metrics.The authors attribute this to global collaborative modeling that preserves individual interaction patterns and cross-view supervision for sparse interactions.
- Contrastive Learning: HCCF achieves larger performance gaps on Amazon data and outperforms MHCN, SGL, and SLRec across datasets.The authors connect this result to handling sparse interactions without relying on random masking that may drop important labels or auxiliary objectives that may misalign with recommendation optimization.
- Ablation Study: Removing hierarchical hypergraph mapping or low-rank dependency encoding degrades accuracy, while low-rank learning also reduces parameter size and may alleviate overfitting.The ablation study evaluates the effects of global hypergraph structure learning and cross-view contrastive self-augmentation.
- Sparsity Analysis: HCCF shows potential for sparse-data robustness across interaction degrees and is more stable than SGL across sparsity levels.The authors attribute this to cooperative supervision between local collaborative relation encoding and global dependency learning.
- Over-Smoothing Analysis: Cross-view hypergraph-guided contrastive learning regularizes embeddings to remain distinct, and HCCF produces higher embedding separation than variants without contrastive learning or hypergraph modeling.A subgraph visualization further shows HCCF distinguishing user groups where the -Hyper variant assigns similar embeddings through smoothing.
5 CONCLUSION
HCCF enhances neural collaborative filtering with hypergraph-guided self-supervised learning, jointly modeling local and global collaborative relationships. Experiments validate its superiority over competitive baselines, while future work proposes incorporating temporal context through dynamic user dependencies.
- HCCF combines hypergraph structure learning with cross-view hypergraph contrastive encoding to improve user representations.
- The framework jointly characterizes local and global collaborative relationships in a shared embedding space.
- Extensive experiments validate HCCF’s superiority over competitive recommendation baselines.
- Future work may model dynamic user dependencies with a time-aware hypergraph embedding function to inject temporal context.