LAGCL4Rec: When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation

Leqi Zheng, Chaokun Wang, Canzhi Chen, Jiajun Zhang, Cheng Wu, Zixin Song, Shannan Yan, Ziyang Liu, Hongwei LiEMNLP (Findings) 2025 · OpenReview

TL;DR

This paper introduces LAGCL4Rec, a novel recommendation framework that leverages Large Language Models (LLMs) to unlock the full potential of user-item interaction data. By intelligently activating both positive and negative feedback signals across data, ranking, and reranking stages, LAGCL4Rec significantly improves recommendation accuracy and efficiency, especially in sparse data scenarios.

What Problem Does This Paper Address?

Traditional recommender systems struggle with several key issues:

  • Sparse User-Item Interactions: It's hard to learn reliable user preferences when there are few interactions recorded.
  • Uniform Treatment of Negative Samples: Contrastive learning methods often treat all negative interactions (items a user didn't interact with) equally, ignoring their varying levels of "difficulty" or semantic dissimilarity, which is a missed opportunity for learning.
  • Discarding Negative Feedback in LLM-based Systems: Modern LLM approaches for recommendation often ignore negative feedback altogether, leading to an unbalanced understanding of user preferences.

These limitations prevent recommender systems from reaching their full potential. The authors aim to systematically activate and leverage all interaction signals, including nuanced negative feedback, to build more accurate and personalized recommendations.

Methods

LAGCL4Rec employs a three-stage "Progressive Activation Pipeline" that integrates LLMs into Graph Contrastive Learning (GCL):

  1. Data-Level Activation:
    • User/Item Profile Generation: LLMs are used to create rich, semantically detailed profiles for users and items, going beyond simple IDs.
  2. Interaction Augmentation: LLMs generate balanced positive and negative interactions. Instead of random negative sampling, they generate "likely negative preferences" by scoring candidate items based on user profiles, creating more informative negative signals.
  3. Rank-Level Activation:
    • LLM-Guided Semantic Grouping: LLMs group users and items into semantically similar clusters. This allows for differentiating negative samples based on their "difficulty" – hard negatives are those within the same semantic group as the positive item, while easy negatives are from different groups.
  4. Difficulty-Aware Contrastive Learning: The framework uses specialized loss functions (LABPR and adaptive InfoNCE) that assign higher importance to hard negatives. This fine-grained supervision helps the model learn more discriminative embeddings.
  5. Rerank-Level Activation:
    • LLM-Based Reasoning: The reranking stage uses LLMs to reason over the augmented historical interactions (both positive and negative).
  6. Chain-of-Thought (CoT) Prompting: A structured CoT process guides the LLM to analyze user preferences, compare candidate items, and produce a final, personalized ranking. This stage is optimized using a binary classification loss.
  7. Score Fusion: The final recommendation score is a weighted combination of traditional collaborative filtering scores and the LLM-generated reranking scores, allowing for a balance between established patterns and nuanced LLM insights.

Datasets

The paper evaluates LAGCL4Rec on four public recommendation datasets:

  • ML-1M
  • Amazon-book
  • Yelp
  • Steam

Standard practices were followed for data splitting and evaluation, using a 7:1:2 ratio for training, validation, and testing.

Key Insights and Ideas

  • Negative Feedback is Crucial and Nuanced: The paper strongly advocates for the importance of negative feedback, not just as a signal of dislike, but as a source of semantically informative signals about user preferences.
  • LLMs as "Preference Activators": LLMs are not just used for feature extraction or generation, but as active agents to "awaken" dormant signals in sparse interaction data and to understand the subtle differences between negative interactions.
  • Semantic Difficulty Matters: Treating negative samples uniformly is suboptimal. Differentiating them based on semantic similarity to positive items allows for more effective contrastive learning and better model convergence.
  • Progressive Activation Pipeline: The multi-stage approach (Data, Rank, Rerank) ensures that LLM capabilities are leveraged throughout the entire recommendation process, leading to synergistic improvements.
  • Efficiency Gains: Despite incorporating LLMs, the proposed rank-level activation strategy significantly reduces total training time by enabling faster convergence, demonstrating that LLM integration doesn't necessarily mean increased computational burden.

Key Results

  • Consistent Performance Improvement: LAGCL4Rec consistently outperforms various baseline models (CF, Sign-aware, LLM-based, GCL) across all four datasets, showing improvements of up to 17.42% in Recall@20 and 15.43% in NDCG@20.
  • Strong Gains on Sparse Data: The method shows particularly substantial gains on datasets like Yelp, which are more categorical and sparse, highlighting its effectiveness in activating sparse signals.
  • Reduced Training Time: LAGCL4Rec introduces minimal per-epoch overhead but significantly reduces total training time (58.7-63.2% reduction) compared to baseline GCL models, due to faster convergence.
  • Component Importance: Ablation studies confirm that each component (Data-Level, Rank-Level, Rerank-Level) is crucial for the overall performance. Removing any component leads to degradation.
  • Hyperparameter Sensitivity: The model shows reasonable sensitivity to key hyperparameters, with optimal settings identified for fusion weight, contrastive weight, and temperature ratios.

Limitations and Future Work

  • Industrial Deployment: The current evaluation is based on public benchmark datasets. A notable limitation is the absence of validation in real-world industrial recommendation systems, which present additional challenges like larger scale, complex interactions, and stricter latency requirements. Future work should focus on adapting and evaluating LAGCL4Rec for industrial deployment.
  • Ethical Considerations: The authors state that their work exclusively uses public, anonymized benchmark datasets and focuses on preference patterns without introducing demographic or social biases. No human subjects were involved.
  • Future Directions: The paper implicitly suggests that further exploration of LLM integration for more sophisticated reasoning and interaction activation in recommendation systems is a promising avenue. The theoretical analysis also points towards the benefits of adaptive learning strategies.

Summary updated

Abstract

A core barrier preventing recommender systems from reaching their full potential lies in the inherent limitations of user-item interaction data: (1) Sparse user-item interactions, making it difficult to learn reliable user preferences; (2) Traditional contrastive learning methods often treat negative samples as equally hard or easy, ignoring the informative semantic difficulty during training. (3) Modern LLM-based recommender systems, on the other hand, discard all negative feedback, leading to unbalanced preference modeling. To address these issues, we propose LAGCL4Rec, a framework leveraging Large Language Models to Activate interactions in Graph Contrastive Learning for Recommendation. Our approach operates through three stages: (i) Data-Level: augmenting sparse interactions with balanced positive and negative samples using LLM-enriched profiles; (ii) Rank-Level: assessing semantic difficulty of negative samples through LLM-based grouping for fine-grained contrastive learning; and (iii) Rerank-Level: reasoning over augmented historical interactions for personalized recommendations. Theoretical analysis proves that LAGCL4Rec achieves effective information utilization with minimal computational overhead. Experiments across multiple benchmarks confirm our method consistently outperforms state-of-the-art baselines.
BibTeX
@inproceedings{zheng2025lagcl,
  author = {Leqi Zheng and Chaokun Wang and Canzhi Chen and Jiajun Zhang and Cheng Wu and Zixin Song and Shannan Yan and Ziyang Liu and Hongwei Li},
  title = {{LAGCL4Rec: When LLMs Activate Interactions Potential in Graph Contrastive Learning for Recommendation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
  url = {https://openreview.net/forum?id=qxzXIrKO6M},
}
Figure