Research map / Graph learning / Graph self-supervised learning
Review-Enhanced Hierarchical Contrastive Learning for Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ReHCL, a graph-based recommender that builds topic and semantic review graphs, combines cross-view graph contrastive learning (with neighbor-based positive sampling) and cross-modal rating-review contrastive learning in a hierarchical objective with BPR.
Problem
Review-enhanced graph-based recommendation fails to fully exploit hidden user/item relations in reviews, filter redundant review information, and model the behavioral association between ratings and reviews.
Approach
Construct topic (LDA-based) and semantic (BERT-based) user/item graphs, contrast them cross-view with neighbor-based intra-view positive pairs, and align LightGCN rating embeddings with review embeddings via a cross-modal contrastive loss with a projection MLP, all jointly trained with the BPR recommendation loss.
Key results
- ReHCL reaches H@10 0.3936 on Instrument, 0.6116 on Music, 0.3885 on Toy, beating the best baseline (HGNR/NARRE) with a 14.68% H@10 improvement on Instrument.
- Review-based baselines outperform rating-only baselines by an average of 5.50% H@10, 11.46% M@10 and 9.49% N@10.
- Ablation: removing CVGCL or CMCL each degrades performance relative to the full hierarchical model, and LightGCN encoder outperforms GCN and NGCF variants.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine CVGCL and CMCL with recommendation training | ReHCL · Instrument, Music, and Toy | HR@10: 0.3432 -> 0.3936; 0.5647 -> 0.6116; 0.3692 -> 0.3885 | Helped |
| combine CVGCL and CMCL with recommendation training | ReHCL · Instrument, Music, and Toy | MRR@10: 0.1492 -> 0.1612; 0.2637 -> 0.2797; 0.1556 -> 0.1833 | Helped |
| combine CVGCL and CMCL with recommendation training | ReHCL · Instrument, Music, and Toy | NDCG@10: 0.2033 -> 0.2155; 0.3346 -> 0.3570; 0.2091 -> 0.2313 | Helped |
| add topic and semantic review graphs to rating-graph modeling (ablation) | ReHCL · Instrument, Music, and Toy | HR@10: 0.2941 -> 0.3936; 0.4897 -> 0.6116; 0.3483 -> 0.3885 | Helped |
| add the rating graph to topic and semantic review graphs (ablation) | ReHCL · Instrument, Music, and Toy | HR@10: 0.3165 -> 0.3936; 0.5121 -> 0.6116; 0.3469 -> 0.3885 | Helped |
| add semantic graphs to the rating graph and topic graphs (ablation) | ReHCL · Instrument, Music, and Toy | HR@10: 0.3242 -> 0.3936; 0.5531 -> 0.6116; 0.3541 -> 0.3885 | Helped |
| add topic graphs to the rating graph and semantic graphs (ablation) | ReHCL · Instrument, Music, and Toy | HR@10: 0.3522 -> 0.3936; 0.5782 -> 0.6116; 0.3657 -> 0.3885 | Helped |
| replace LightGCN with GCN (ablation) | ReHCL · Instrument and Music | HR@10: 0.3578 -> 0.3936; 0.5601 -> 0.6116 | Helped |
| replace LightGCN with NGCF (ablation) | ReHCL · Instrument and Music | HR@10: 0.3690 -> 0.3936; 0.5831 -> 0.6116 | Helped |
| use embedding-based augmentation instead of no augmentation (ablation) | ReHCL · Instrument, Music, and Toy | HR@10: 0.3873 -> 0.3936; 0.6031 -> 0.6116; 0.3788 -> 0.3885 | Helped |
| use embedding-based augmentation instead of edge dropout (ablation) | ReHCL · Instrument and Toy | HR@10: 0.3898 -> 0.3936; 0.3821 -> 0.3885 | Helped |
| use neighbor-based positive sampling instead of InfoNCE (ablation) | ReHCL · Instrument, Music, and Toy | HR@10: 0.3908 -> 0.3936; 0.6042 -> 0.6116; 0.3806 -> 0.3885 | Helped |
Limitations
- Experiments limited to three Amazon datasets with small-to-medium scale.
- Topic number and layer number sensitivity; over-smoothing beyond 6 layers.
- Relies on fixed LDA and BERT pipelines for graph construction without end-to-end learning.
Experimental setup
Datasets: Amazon Instrument Amazon Music Amazon Toy
Benchmarks: Amazon Instrument Amazon Music Amazon Toy
Metrics: HR@10 MRR@10 NDCG@10
Baselines: GC-MC GCN NGCF LightGCN SGL TopicMF ALFM DeepCoNN NARRE SSG HGNR RGCL
Builds on: SGL LightGCN InfoNCE CLIP RGCL
Closest papers by meaning
- SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.88)
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.86)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.86)
- Contrastive Representation for Interactive Recommendation (AAAI 2025 · similarity 0.85)
- SHARP-Distill: A 68× Faster Recommender System with Hypergraph Neural Networks and Language Models (ICML 2025 · similarity 0.85)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.