[Paper Review] Disentangled Graph Contrastive Learning for Review-based Recommendation

Yuyang Ren, Haonan Zhang+6 more|arXiv (Cornell University)|Sep 4, 2022

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes DGCLR, a disentangled graph contrastive learning framework that models user-item interactions through distinct latent factors (e.g., quality, price, appearance) using review text and user-item interaction graphs. By learning factor-specific representations via a factorized message passing mechanism and contrastive learning, DGCLR improves recommendation accuracy and interpretability, outperforming state-of-the-art methods on five benchmark datasets.

ABSTRACT

User review data is helpful in alleviating the data sparsity problem in many recommender systems. In review-based recommendation methods, review data is considered as auxiliary information that can improve the quality of learned user/item or interaction representations for the user rating prediction task. However, these methods usually model user-item interactions in a holistic manner and neglect the entanglement of the latent factors behind them, e.g., price, quality, or appearance, resulting in suboptimal representations and reducing interpretability. In this paper, we propose a Disentangled Graph Contrastive Learning framework for Review-based recommendation (DGCLR), to separately model the user-item interactions based on different latent factors through the textual review data. To this end, we first model the distributions of interactions over latent factors from both semantic information in review data and structural information in user-item graph data, forming several factor graphs. Then a factorized message passing mechanism is designed to learn disentangled user/item representations on the factor graphs, which enable us to further characterize the interactions and adaptively combine the predicted ratings from multiple factors via a devised attention mechanism. Finally, we set two factor-wise contrastive learning objectives to alleviate the sparsity issue and model the user/item and interaction features pertinent to each factor more accurately. Empirical results over five benchmark datasets validate the superiority of DGCLR over the state-of-the-art methods. Further analysis is offered to interpret the learned intent factors and rating prediction in DGCLR.

Motivation & Objective

• To address the limitation of holistic modeling in existing review-based recommendation methods, which entangle multiple latent factors like price, quality, and appearance in user-item representations.
• To improve recommendation performance and interpretability by explicitly disentangling user-item interactions into distinct latent factors using textual reviews and user-item graph structures.
• To learn disentangled user and item representations that capture fine-grained preferences per factor, enabling more accurate and explainable rating predictions.
• To alleviate data sparsity by introducing factor-wise contrastive learning objectives that enhance the discriminative power of representations for each latent factor.

Proposed method

• The framework constructs multiple factor graphs by modeling the distribution of interactions across latent factors using both semantic signals from reviews and structural signals from the user-item graph.
• A factorized message passing mechanism is designed to learn disentangled user and item representations on each factor graph, capturing factor-specific preferences.
• An attention-based interaction module adaptively combines predictions from multiple factors to generate final rating scores, with attention weights indicating the dominant factor for each interaction.
• Two factor-wise contrastive learning objectives are introduced: one for user/item representations and one for interaction representations, to enhance disentanglement and robustness.
• The model is trained end-to-end using a multi-task loss combining rating prediction and contrastive learning objectives, with hyperparameters tuned for optimal performance.

Experimental results

Research questions

• RQ1 Can disentangling latent factors (e.g., quality, price, appearance) in user-item interactions improve recommendation performance and interpretability?
• RQ2 How can review text and user-item graph structures be jointly leveraged to identify and model distinct latent factors behind user ratings?
• RQ3 Does factor-wise contrastive learning improve the discriminative quality and disentanglement of learned representations?
• RQ4 How do different numbers of factors and message passing layers affect model performance and generalization?

Key findings

• DGCLR achieves state-of-the-art performance on five benchmark datasets, outperforming existing methods in rating prediction accuracy.
• The model shows improved interpretability: factor-specific reviews consistently reflect coherent semantic themes such as 'demand-supply match', 'quality', 'integrity', and 'cost performance'.
• Performance degrades with deeper message passing when factors are not disentangled (K=1), but improves significantly when K≥2, indicating that disentanglement enables effective higher-order message passing.
• The optimal contrastive loss weights (λ₁, λ₂) fall in the range [0.3, 0.7], and the model is robust to hyperparameter tuning within this range.
• The attention mechanism successfully identifies dominant factors for individual predictions, with high attention weights (>0.5) aligning with semantically coherent review content.
• The model achieves accurate rating predictions (MSE < 0.2) by focusing on the most relevant factor per interaction, demonstrating both precision and interpretability.