PaperlayerView exact revision on arXiv

Source-linked AI summary

A Review-aware Graph Contrastive Learning Framework for Recommendation

Jie Shuai, Kun Zhang, Le Wu, Peijie Sun, Richang Hong, Meng Wang, Yong Li

arXiv:2204.12063v4cs.IR

TL;DR

Review-based recommendation must better exploit user-item graphs with review-featured edges and provide self-supervised signals despite limited user behaviors. RGCL constructs such a graph, uses review-aware graph learning, and adds Node and Edge Discrimination tasks; experiments over five datasets report superiority over state-of-the-art baselines.

  • Problem

    Existing review-based methods underuse the user-item graph with review-derived edge features and lack self-supervised signals for limited user behaviors.

  • Method

    RGCL constructs a user-item graph with rating and review-semantic edge features, then uses review-aware graph learning plus Node Discrimination and Edge Discrimination tasks.

  • Results

    RGCL is reported effective and superior to state-of-the-art review-based baselines in experiments over five datasets.

  • Takeaways & Limitations

    Review-aware graph learning and the two contrastive tasks provide a framework for using review information in embedding learning and interaction modeling.

  • Takeaways & Limitations

    The method uses pre-encoded reviews that embody ratings to reduce training time and memory without decreasing recommendation performance.

Abstract

from arXiv · show

Most modern recommender systems predict users preferences with two components: user and item embedding learning, followed by the user-item interaction modeling. By utilizing the auxiliary review information accompanied with user ratings, many of the existing review-based recommendation models enriched user/item embedding learning ability with historical reviews or better modeled user-item interactions with the help of available user-item target reviews. Though significant progress has been made, we argue that current solutions for review-based recommendation suffer from two drawbacks. First, as review-based recommendation can be naturally formed as a user-item bipartite graph with edge features from corresponding user-item reviews, how to better exploit this unique graph structure for recommendation? Second, while most current models suffer from limited user behaviors, can we exploit the unique self-supervised signals in the review-aware graph to guide two recommendation components better? To this end, in this paper, we propose a novel Review-aware Graph Contrastive Learning (RGCL) framework for review-based recommendation. Specifically, we first construct a review-aware user-item graph with feature-enhanced edges from reviews, where each edge feature is composed of both the user-item rating and the corresponding review semantics. This graph with feature-enhanced edges can help attentively learn each neighbor node weight for user and item representation learning. After that, we design two additional contrastive learning tasks (i.e., Node Discrimination and Edge Discrimination) to provide self-supervised signals for the two components in recommendation process. Finally, extensive experiments over five benchmark datasets demonstrate the superiority of our proposed RGCL compared to the state-of-the-art baselines.

1 INTRODUCTION

Review-based recommendation uses reviews either to enrich user and item representations or to model specific user-item interactions, but existing methods underuse the review-aware graph structure and lack self-supervised signals for limited behaviors. RGCL addresses these challenges with review-aware graph learning and Node Discrimination and Edge Discrimination tasks, with experiments on five datasets showing superiority over state-of-the-art baselines.

  • Existing review-based recommendation: Review-based recommendation enriches user and item representations with historical reviews or models interactions using target reviews.These correspond to the two main recommendation components: embedding learning and user-item interaction modeling.
  • Challenges: Current methods face challenges in exploiting the user-item bipartite graph with review-derived edge features and developing self-supervised signals under limited user behaviors.The paper identifies both graph-structure utilization and self-supervision as unresolved challenges.
  • Proposed framework: RGCL integrates reviews into graph learning so they determine neighbor-node and review impacts on central user and item representations.Its Review-aware Graph module exploits the graph structure formed by user-item edges with review features.
  • Proposed framework: Node Discrimination and Edge Discrimination provide additional self-supervised optimization targets for embedding learning and interaction modeling.The two tasks are designed to use review information in graph edges for the two recommendation components.
  • Evaluation: Experiments and detailed analysis over five datasets show RGCL is effective and superior to state-of-the-art review-based baselines.The contribution summary likewise reports extensive experiments on five real-world datasets.

2 RELATED WORK

Related work mainly uses reviews for historical representation learning or target interaction modeling, while graph-based and contrastive recommendation methods provide relevant foundations. RGCL extends these directions by constructing a review-featured user-item graph and adding self-supervised tasks for limited user behaviors.

  • Historical reviews utilization: Historical-review methods improve user and item embeddings with topic models, text encoders, and attention mechanisms.Examples include LDA, CTR, TIM, DeepCoNN, CARL, DAML, and AHN.
  • Graph-based recommendation: Graph-based review recommendation models user-item bipartite structure while introducing review signals into graph learning.RMG combines review features with attention-based graph convolution for node embeddings.
  • RGCL positioning: RGCL advances related graph methods by constructing a user-item graph with edge features from corresponding reviews.This design is intended for user preference modeling and item recommendation.
  • Target reviews utilization: Target-review methods incorporate reviews into user-item interaction modeling, but target reviews are unavailable at inference for unobserved user-item pairs.This motivates methods designed to address the mismatch between training-time target reviews and inference-time availability.
  • Contrastive learning: Existing review-based methods treat reviews as additional information and suffer from limited user behaviors, motivating self-supervised signals from two contrastive-learning tasks.The paper positions its CL tasks as a response to this limitation.
  • Contrastive learning: Contrastive learning has also been applied to graph representation learning and graph-based recommendation for representation or interaction modeling.Prior paradigms include DGI, BiGI, and other graph contrastive methods.

3 PROBLEM DEFINITION

The paper formulates review-based recommendation as rating prediction on a user-item bipartite graph whose observed edges contain both ratings and review-semantic features.

  • Data representation: The data contain users U, items V, a rating matrix R, and a review-feature tensor E obtained from pre-trained review representations.Users and items are indexed separately, while each review is encoded as a fixed-size vector.
  • Featured-edge graph: Each observed user-item edge contains a detailed rating value and the semantic vector of its corresponding review.Together, the rating matrix and review tensor define the featured-edge interaction set.
  • Prediction task: The recommendation task is to predict the final rating matrix from the user-item graph with featured edges.The graph is represented as G = < U ∪ V, E >.

4 REVIEW-AWARE GRAPH CONTRASTIVE LEARNING

RGCL integrates ratings and review semantics into user-item graph learning, then adds contrastive tasks for node representations and interaction modeling.

  • Framework overview: RGCL has a Review-aware Graph Learning Module and a Graph Contrastive Learning Module targeting representation learning and interaction modeling.The framework incorporates reviews into graph-based user preference modeling and applies contrastive learning from two perspectives.
  • Review-aware graph learning: The model initializes user and item embeddings with rating-specific edge types and frozen BERT-Whitening review representations.Each review is encoded as a fixed-size vector, while rating scores distinguish edge semantics during graph learning.
  • Review-aware graph learning: Review-aware message passing uses rating and review features to re-weight neighbor and review influences on each central node.The review feature helps model detailed preferences and item attributes while fine-tuning message impacts.
  • Interaction modeling: Messages are aggregated across layers to produce final user and item embeddings, which an MLP transforms into user-item interaction features for rating prediction.The learned interaction features are mapped to predicted ratings through the interaction-modeling component.
  • Graph contrastive learning: Node Discrimination uses node dropping and contrasting embeddings from two subgraphs to improve user and item representation learning.Positive pairs represent the same node across subgraphs, while embeddings from different nodes provide negative examples.
  • Graph contrastive learning: Edge Discrimination aligns each learned interaction feature with its corresponding review representation while contrasting it against sampled reviews.This task uses the interaction feature as an anchor, the target review as a positive sample, and another training-set review as a negative sample.

5.1 Experimental Settings

The experiments use five review-based recommendation datasets, MSE evaluation, repeated runs, and baselines spanning rating-based, review-based, graph-based, and graph-review methods.

  • Datasets: Five benchmark datasets include four Amazon 5-core domains and the Yelp Business Rating Prediction Challenge 2013 dataset.The Amazon domains are Digital Music, Toys and Games, Clothing, and CDs and Vinyl; Yelp covers restaurant reviews in metropolitan Phoenix.
  • Evaluation: Performance is evaluated with MSE, and each experiment is repeated five times with mean(±std) reported.Relative improvements above 1% are regarded as significant for review-based recommendation.
  • Baselines: The comparison includes traditional rating-based, historical-review, target-review, graph-based, and graph-review-fusing baselines.Baselines include SVD, NCF, DeepCoNN, NARRE, DAML, SDNet, TransNets, GC-MC, RMG, and SSG.
  • Reporting: Table 2 reports MSE results across the five datasets, with best results bolded and relative improvements measured against the best underlined baselines.Results are reported as mean(±std) across five random runs.
  • Implementation: Reviews are encoded with BERT-Whitening into fixed-size vectors that remain unchanged during training.The final RG architecture uses one message-passing layer and embedding size d=64.

5.2 Performance Evaluation

RGCL achieves the best performance across five datasets and remains strongest across user groups with different interaction densities. Its graph learning and contrastive tasks are associated with improved use of reviews under sparse behavior.

  • Overall Performance: Target reviews improve performance because they are highly relevant to the corresponding users and items.Review-based baselines also show the effectiveness of incorporating review information.
  • Overall Performance: Graph-based baselines achieve the best performance among the baselines, supporting graph learning for modeling higher-order user-item signals.The authors speculate that SSG performs worse because it ignores collaborative filtering signals among users and items.
  • Overall Performance: RGCL achieves the best performance across all five datasets compared with the review-based and graph-based baselines.The authors attribute this to review-aware graph learning together with node and edge contrastive learning.
  • Data Sparsity: Node and edge contrastive learning improve review utilization under data sparsity, with edge discrimination focusing on consistency between interaction and review representations.The user-group analysis compares RG, RG+ND, RG+ED, and RGCL across groups ranked by interaction counts.
  • Data Sparsity: RGCL has the best performance across different data-sparsity settings by combining historical and target review utilization with contrastive learning.The framework applies the contrastive tasks to user/item embedding and interaction modeling.

5.3 The Effectiveness of RG

Ablation results show that RG benefits from both review-based neighbor weighting and review features, while one message-passing layer gives the lowest reported MSE among the tested depths.

  • RG Ablation: Both RG(w/o review) and RG(w/o weight) outperform GC-MC and RGAT, indicating that RG better explores review information.The ablations isolate review use in neighbor weighting versus review use as an impact factor.
  • RG Ablation: RG(w/o review) performs better than RG(w/o weight), showing that reviews are critical for determining neighbor-impact weights.The comparison supports the importance of review-derived weighting in the graph operation.
  • RG Ablation: Integrating both review operations enables RG to combine historical-review utilization with target-review utilization.The RG design is intended to exploit both review roles within graph learning.
  • Layer Depth: RG achieves the smallest MSE with one message-passing layer among the tested one-, two-, and three-layer settings.Two layers produce a slight MSE decrease relative to one layer, while three layers produce a larger performance reduction.

5.4 The Effectiveness of CL tasks

The experiments examine how the ED and ND contrastive-learning tasks affect RGCL and whether they generalize across recommendation models. Results show that task strength matters, both tasks benefit different backbones, and RGCL learns more consistent interaction features.

  • 5.4.1 Parameter Sensitive Experiments: ND strength also depends on the dataset, with best β values of 0.2 for Digital Music and 0.6 for Clothing.The paper attributes this difference to dataset-specific data size, sparsity, and overfitting tendencies.
  • 5.4.2 Generalization of CL Tasks: Both ED and ND improve the performance of the evaluated NARRE and GC-MC backbones, supporting generalization across review-based recommendation models.The experiments impose ED on NARRE and both tasks on GC-MC.
  • 5.4.2 Generalization of CL Tasks: GC-MC gains 3.5%, 2.0%, and 1.7% when both ED and ND are incorporated across three datasets.The improvement exceeds that observed for NARRE, and the authors connect ED with accessing information during complex interaction modeling.
  • 5.4.3 Quantitative Analysis about the Learned Interaction Features: RGCL produces the best consistency between learned user-item interaction features and corresponding review representations among the compared methods.The analysis estimates KL-divergence with MINE, where a larger value indicates higher dependence; RGCL also outperforms TransNets and SDNet without CL tasks.

6 CONCLUSION

The paper proposes RGCL to exploit review-enhanced user-item graph structure for preference modeling and rating prediction. It combines review-aware graph learning with ED and ND contrastive tasks, and reports effectiveness across five benchmark datasets while identifying future work on review representations and fusion.

  • 6 CONCLUSION: RGCL models user preferences and predicts ratings by exploiting a user-item bipartite graph whose edges contain ratings and review information.The framework is designed to enhance both user/item embedding learning and user-item interaction modeling.
  • 6 CONCLUSION: The review-aware graph learning module uses reviews to fine-tune the influence of corresponding neighbors and the reviews themselves.This module incorporates review information into user and item embedding learning.
  • 6 CONCLUSION: ED and ND constrain RGCL to improve node embeddings and interaction modeling.These are the two additional contrastive-learning tasks introduced in the framework.
  • 6 CONCLUSION: Experiments on five benchmark recommendation datasets demonstrate RGCL’s effectiveness and superiority.
  • 6 CONCLUSION: Future work includes advanced review embedding methods and improved fusion and alignment for complex review-user or review-item interactions.
Loading 2204.12063v4…