Research map / Graph learning / Graph self-supervised learning
Semantic Enhanced Heterogeneous Hypergraph Network for Collaborative Filtering
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 SEHHN, a heterogeneous hypergraph network that injects LLM-derived review semantics and LLM-inferred item categories into collaborative filtering via a graph autoencoder, avoiding alignment degradation from inconsistent text semantics.
Problem
LLM-enhanced collaborative filtering suffers from semantic inconsistency between user-item and item-item text pairs, which degrades alignment of textual and collaborative embeddings in recommender systems.
Approach
Instead of directly aligning LLM text embeddings with CF embeddings, it enriches the CF relations themselves: review nodes and LLM-classified category nodes are placed in a heterogeneous hypergraph with four hyperedge types, and a graph autoencoder reconstructs review semantics from collaborative signals.
Key results
- On Book, SEHHN reaches Recall@20 0.1677 and MRR@20 0.1630, versus 0.1608 and 0.1609 for RLMRec (best baseline).
- On Steam, SEHHN reaches Recall@10 0.0965 versus 0.0891 for RLMRec.
- On Yelp, SEHHN reaches Recall@10 0.0836 versus 0.0812 for RLMRec; ablations show removing reviews or categories lowers performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SEHHN instead of the strongest baseline | Collaborative filtering · Book | Recall@20: 0.1608 -> 0.1677 | Helped |
| use SEHHN instead of the strongest baseline | Collaborative filtering · Yelp | Recall@20: 0.1326 -> 0.1360 | Helped |
| use SEHHN instead of the strongest baseline | Collaborative filtering · Steam | Recall@20: 0.1450 -> 0.1512 | Helped |
| remove review information (ablation) | SEHHN · Book and Yelp | recommendation performance | Hurt |
| remove category information (ablation) | SEHHN · Book and Yelp | recommendation performance | Hurt |
| replace LLM review embeddings with random embeddings (ablation) | SEHHN · Book and Yelp | recommendation performance | Hurt |
| remove the graph-autoencoder reconstruction task (ablation) | SEHHN · Book and Yelp | recommendation performance | Hurt |
| replace the graph encoder with a linear transformation (ablation) | SEHHN · Book and Yelp | recommendation performance | Hurt |
| replace the graph decoder with direct reconstruction from review embeddings (ablation) | SEHHN · Book and Yelp | recommendation performance | Hurt |
| replace the default embedder with BGE-large (ablation) | SEHHN · Yelp | Recall@20: 0.1351 -> 0.1322 | Hurt |
| replace the default embedder with Bloom-1b1 (ablation) | SEHHN · Yelp | Recall@20: 0.1351 -> 0.1316 | Hurt |
| use heterogeneous aggregation rather than no aggregation (L = 0) (ablation) | SEHHN · Book and Yelp | recommendation performance | Helped |
Surprising findings
- Replacing LLM review embeddings with random embeddings (SEHHN-RE) reduces performance, but the gap from an LLM-free variant is reported as the value of LLM semantics rather than as a surprise.
Limitations
- Relies on proprietary GPT-3.5-turbo and text-embedding-ada-002 via OpenAI API, with open-source alternatives only tested for embeddings.
- Over-smoothing with too many GNN layers and sensitivity to reconstruction ratio are reported.
- Reviews are needed for training interactions, so the approach depends on available review text.
Experimental setup
Datasets: Amazon Book Yelp Steam
Benchmarks: Book Yelp Steam
Metrics: Recall@10 Recall@20 MRR@10 MRR@20
Baselines: GCCF LightGCN SGL SimGCL DCCF AutoCF RLMRec
Builds on: LightGCN RLMRec SSLRec GAT BPR
Closest papers by meaning
- SHARP-Distill: A 68× Faster Recommender System with Hypergraph Neural Networks and Language Models (ICML 2025 · similarity 0.89)
- Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering (ICML 2025 · similarity 0.87)
- When Hypergraph Meets Heterophily: New Benchmark Datasets and Baseline (AAAI 2025 · similarity 0.87)
- Heterogeneous Graph Neural Network on Semantic Tree (AAAI 2025 · similarity 0.87)
- Semantic Convergence: Harmonizing Recommender Systems via Two-Stage Alignment and Behavioral Semantic Tokenization (AAAI 2025 · similarity 0.86)
- FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.86)
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.