PaperlayerView exact revision on arXiv

Source-linked AI summary

Disentangled Contrastive Collaborative Filtering

Xubin Ren, Lianghao Xia, Jiashu Zhao, Dawei Yin, Chao Huang

arXiv:2305.02759v4cs.IRcs.AI

TL;DR

GCL-based recommender systems face entangled intent factors and non-adaptive, noise-sensitive augmentation, motivating a more robust approach. DCCF combines global-context disentangled representations with adaptive interaction masking and reports superior performance across public datasets and more than 10 existing solutions.

  • Problem

    Existing GCL-based recommenders overlook diverse latent intents and use augmentation vulnerable to noisy information, while coarse-grained preference modeling can yield suboptimal representations.

  • Method

    DCCF learns global-context multi-intent representations and uses parameterized mask generators with intent-aware contrastive learning for adaptive interaction augmentation.

  • Results

    DCCF achieves superior recommendation performance compared to more than 10 existing solutions across extensive experiments on different recommendation datasets.

  • Takeaways & Limitations

    Disentangled self-supervision with adaptive augmentation provides a framework for reducing augmentation-induced noise while modeling finer-grained interaction factors.

  • Takeaways & Limitations

    The related-work discussion notes that coarse-grained user-preference modeling can produce suboptimal recommendation representations.

Abstract

from arXiv · show

Recent studies show that graph neural networks (GNNs) are prevalent to model high-order relationships for collaborative filtering (CF). Towards this research line, graph contrastive learning (GCL) has exhibited powerful performance in addressing the supervision label shortage issue by learning augmented user and item representations. While many of them show their effectiveness, two key questions still remain unexplored: i) Most existing GCL-based CF models are still limited by ignoring the fact that user-item interaction behaviors are often driven by diverse latent intent factors (e.g., shopping for family party, preferred color or brand of products); ii) Their introduced non-adaptive augmentation techniques are vulnerable to noisy information, which raises concerns about the model's robustness and the risk of incorporating misleading self-supervised signals. In light of these limitations, we propose a Disentangled Contrastive Collaborative Filtering framework (DCCF) to realize intent disentanglement with self-supervised augmentation in an adaptive fashion. With the learned disentangled representations with global context, our DCCF is able to not only distill finer-grained latent factors from the entangled self-supervision signals but also alleviate the augmentation-induced noise. Finally, the cross-view contrastive learning task is introduced to enable adaptive augmentation with our parameterized interaction mask generator. Experiments on various public datasets demonstrate the superiority of our method compared to existing solutions. Our model implementation is released at the link https://github.com/HKUDS/DCCF.

1 INTRODUCTION

Graph contrastive recommender systems address sparse supervision but overlook entangled intent factors and noise-sensitive augmentation. DCCF addresses both through disentangled representations and adaptive, parameterized interaction masking.

  • Motivation: Existing GCL recommender systems use contrastive views to provide self-supervision for sparse user-item interactions.They learn invariant user and item representations by maximizing agreement between augmented views.
  • Limitations: Current methods overlook diverse, entangled intent factors, limiting their ability to capture fine-grained interaction patterns and accurate intent-aware self-supervision.Examples include shopping for a family party or preferring particular clothing characteristics.
  • Limitations: Random augmentation can damage semantic relationships and important inter-community connections, producing noisy self-supervised signals.Dropping hub nodes is identified as one way stochastic augmentation can harm the relation graph.
  • DCCF: DCCF learns multi-intent representations using global dependencies, intent-aware information passing, and aggregation between patch-level nodes and global intent prototypes.This design aims to distill useful environment-invariant structural information while reducing severe noise in self-supervision.
  • DCCF: Parameterized edge mask generators make contrastive augmentation adaptive to the importance of individual user-item interactions.The generators capture implicit user-item relationships and produce an adaptive graph structure masker.
  • Evidence: DCCF is evaluated against more than 10 existing solutions, with experiments reported as demonstrating superior recommendation performance.The introduction also states that ablation, robustness, and interpretability studies support the disentangled adaptive augmentation.

2 RELATED WORK

Related recommender research combines GNN-based collaborative modeling with disentangled representations, but coarse-grained preference modeling remains a limitation.

  • GNNs-based Recommender Systems: GNN-based recommenders use recursive message passing to model high-order collaborative relations in user-item graphs.Examples include LightGCN, LR-GCCF, and HGCF, alongside graph-enhanced attention mechanisms in other recommendation settings.
  • Recommendation with Disentangled Representations: Existing recommendation approaches learn disentangled user latent intents from implicit feedback using methods including variational autoencoders.The related-work discussion places disentangled representation learning among established recommendation directions.
  • Limitations: Coarse-grained user-preference modeling can produce suboptimal representations for recommendation.This is identified as a limitation in the related-work discussion.

3 METHODOLOGY

DCCF models user-item preferences through latent intents and global context, then uses learnable interaction masks to generate adaptive contrastive views. This design aims to disentangle interaction factors while reducing noise from fixed augmentation.

  • 3.1.1 Modeling Latent Intent Factors: DCCF represents user-item preference as interactions between latent user and item intents, with prediction based on their encoded representations.The model approximates the expected forecasting function over intent distributions using the corresponding expected intent representations.
  • 3.1.2 Multi-Intent Representation with Global Context: The model combines local graph propagation with global intent prototypes to model collaborative dependencies beyond direct user-item connections.Normalized adjacency propagation aggregates neighboring representations, while global prototypes provide intent-aware information aggregation.
  • 3.1.2 Multi-Intent Representation with Global Context: Intent-aware aggregation and residual message passing preserve disentangled user and item representations across graph layers.DCCF defines separate global intent prototypes and integrates their context with graph-based collaborative signals.
  • 3.2 Adaptive Augmentation: The augmentation integrates local collaborative signals with global disentangled relationships to produce multiple contrastive representation views.These views include adaptive local, disentangled global, and adaptive augmented representations used in contrastive learning.
  • 3.2 Adaptive Augmentation: DCCF learns layer-specific interaction masks from interaction importance to construct adaptive relation matrices for graph propagation.Mask values reflect the importance of observed interactions, and the learned relation matrix is normalized before message passing.

4 EVALUATION

The evaluation examines DCCF through five research questions covering comparative performance, component benefits, sparsity, latent-intent count, and training efficiency.

  • The experiments assess whether DCCF outperforms recommendation baselines across different experimental settings.
  • The evaluation studies whether DCCF’s key components improve representation learning and performance.
  • The experiments test whether disentangled self-supervised signals alleviate data sparsity issues.
  • Additional questions examine the impact of latent-intent count and DCCF’s training efficiency.

4.1 Experimental Settings

The experiments use three public recommendation datasets, all-rank evaluation with Recall@N and NDCG@N, diverse baseline families, and standardized implementation settings.

  • Datasets: The evaluation uses Gowalla, Amazon-book, and Tmall datasets representing check-ins, book ratings, and online purchases.
  • Evaluation Protocol: All methods are evaluated over all items using Recall@N and NDCG@N to measure recommendation accuracy.
  • Baseline Methods: Baselines cover factorization, autoencoder, graph neural network, disentangled multi-intent, and self-supervised recommendation methods.
  • Baseline Methods: Compared self-supervised methods include SLRec, SGL-ED/ND, HCCF, and LightGCL.
  • Implementation Settings: DCCF uses Adam with learning rate 1e-3, selects K from {32, 64, 128, 256}, and defaults to K=128.
  • Implementation Settings: Embedding dimensionality and batch size are fixed at 32 and 10240 for compared methods, while graph models use one to three propagation layers.

4.2 Performance Comparison (RQ1)

DCCF consistently outperforms the evaluated baselines across all three datasets, including self-supervised and disentangled recommenders, while its evaluation reports Recall and NDCG.

  • Overall Results: DCCF consistently outperforms all baselines on all three datasets.
  • Evaluation Measures: Table 2 compares methods on different datasets using Recall and NDCG.
  • Compared with Self-Supervised Methods: DCCF achieves superior performance over augmentation-based methods such as SGL and HCCF, and reports state-of-the-art performance across all datasets.
  • Compared with Disentangled Methods: DCCF also improves over disentangled recommenders including DGCF, DisenGCN, and CDR.
  • Sparsity Analysis: Figure 2 compares recommendation performance across user and item groups with different data sparsity on Gowalla.

4.3 Ablation Study (RQ2)

The ablation study evaluates disentangled multi-intent encoding, parameterized adaptive masking, and self-supervised learning as components of DCCF.

  • Ablation Design: The study removes or varies DCCF components to assess their individual contributions to performance.
  • Disentangled Multi-intent Encoding: Removing disentangled multi-intent encoding creates the -Disen variant for measuring the contribution of multi-intent representation encoding.
  • Embedding Smoothness: Table 4 measures embedding smoothness on Amazon-book and Tmall using MAD, where smaller values indicate more obvious over-smoothing.
  • Parameterized Adaptive Masking: Removing local relation learning or intent-based graph structure learning produces -LocalR and -DisenR variants, both of which degrade performance.
  • Self-Supervised Learning: The study also evaluates disentangled contrastive learning by removing agreements between self-supervised objectives.

4.4 In-Depth Analysis of DCCF (RQ3 & RQ4)

DCCF remains effective across interaction sparsity levels and latent-intent settings, while excessive intent granularity can introduce redundancy and noise. Its computational cost is also evaluated against other disentangled recommenders.

  • Robustness to Data Sparsity: DCCF consistently outperforms LightGCN, DGCL, and DGCF for both inactive and active users, indicating stronger performance under sparse data.It also maintains an apparent advantage on low-degree items through self-supervised information that improves long-tail interaction modeling.
  • Impact of the Number of Intent Prototypes: Increasing latent intents generally improves performance, but gains become limited from 128 to 256 and degrade on Tmall.The evaluated intent counts are 32, 64, 128, and 256.
  • Impact of the Number of Intent Prototypes: At 128 intents, some prototypes already cluster; further increases create redundant, overly fine-grained factors and introduce representation-learning noise.This clustering-based explanation is obtained by visualizing intent prototypes in 2D with t-SNE.
  • Robustness to Over-Smoothing: DCCF’s over-smoothing behavior is assessed using Mean Average Distance on normalized user and item embeddings, compared with DCCF-CL, DGCL, DisenGCN, and LightGCN.DCCF-CL disables the cross-view contrastive learning module.
  • Training Efficiency: Table 5 compares per-epoch training time in seconds for DCCF, DGCF, and DisenHAN on Gowalla, Amazon, and Tmall.The supplied materials identify the comparison scope but do not provide the table’s numerical values.

4.5 Model Training Efficiency Study (RQ5)

DCCF is evaluated for training efficiency against disentangled recommender systems on Gowalla, Amazon, and Tmall. The study reports comparable training efficiency while identifying recursive routing and graph attention as costs in competing methods.

  • Training Efficiency: DCCF achieves comparable training efficiency to DGCF and DisenHAN across the evaluated datasets.The comparison uses training computational cost measured per epoch.
  • Training Efficiency: DGCF incurs extra processing time from recursively routing information to obtain intent-relevant weights.Its user embedding is split into intent-aware vectors, but the routing mechanism requires multiple propagation iterations.
  • Training Efficiency: DisenHAN has higher computational cost because its graph attention network must compute attention weights.

4.6 Case Study

The case studies show that DCCF captures global intent-aware relations beyond local graph connectivity and assigns adaptive interaction weights that reflect user preferences.

  • Global Intent-aware Semantic Dependency: DCCF identifies non-locally connected users u1155 and u32856 as similar because their interacted-item categories substantially overlap.The overlapping categories are 29, 36, and 69, and the similarity appears in intent-prototype-specific user weights.
  • Intent-aware Adaptive Augmentation: The learned connectivity weight is consistent with preference degree: higher DCCF weights indicate stronger user preference.This relation is illustrated over the user-item interaction matrix.
  • Intent-aware Adaptive Augmentation: For user u22567, DCCF groups interacted items by category and learns user-item connectivity weights through adaptive augmentation.The case examines categories including 1024 and 645.

5 CONCLUSION

DCCF combines disentangled contrastive learning with adaptive interaction augmentation to model latent intents in recommendation. Experiments across recommendation datasets validate the method’s effectiveness, while future work targets causal analysis and personalized augmentation.

  • Conclusion: DCCF explores latent factors behind implicit interaction intents and uses learned disentangled dependencies for adaptive interaction augmentation.Its graph structure learning layer constructs augmented intent-aware graph structures, followed by intent-aware contrastive learning.
  • Conclusion: Extensive experiments across different recommendation datasets validate the effectiveness of the proposed model.The conclusion states this result without reporting a specific metric.
  • Conclusion: Future directions include combining disentangled representation learning with causal analysis and tailoring augmentation to user characteristics.These directions address noisy-interaction bias and customized graph perturbation operations.
Loading 2305.02759v4…