Research map / Graph learning / Knowledge and temporal graphs
Feature-Structure Adaptive Completion Graph Neural Network for Cold-start 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 FS-GNN, a cold-start recommender that jointly completes missing node features (via LLM-generated features refined by a graph auto-encoder) and missing structure (three-channel PPR and kNN completion) before LightGCN-based rating prediction.
Problem
Cold-start recommendation where new users/items lack both interactions (structure) and feature information, which existing meta-learning and GNN methods handle only partially.
Approach
Treats cold-start as joint feature and structure missingness in a user-item bipartite graph: LLM-based feature completion refined by a GAT/GCN graph auto-encoder, plus a three-channel structure completion (PPR for U-I, feature-similarity kNN for U-U and I-I), trained end-to-end with an adaptive combination of feature, structure, and prediction losses.
Key results
- On user cold-start, FS-GNN improves over the best baseline AGNN by 6.9%, 5.5%, and 6.4% on MovieLens100K, MovieLens1M, and Yelp respectively.
- On item cold-start, FS-GNN outperforms AGNN by 6.0%, 8.8%, and 9.3% on the same datasets.
- In warm-start, FS-GNN is sub-optimal, slightly behind AGNN and ColdGPT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feature completion and three-channel structure completion to LightGCN | FS-GNN · MovieLens100K UCS | MAE: 0.7802 -> 0.7735 | Helped |
| Add feature completion and three-channel structure completion to LightGCN | FS-GNN · MovieLens100K ICS | MAE: 0.8576 -> 0.8064 | Helped |
| Add feature completion and three-channel structure completion to LightGCN | FS-GNN · MovieLens1M UCS | MAE: 0.7957 -> 0.7653 | Helped |
| Add feature completion and three-channel structure completion to LightGCN | FS-GNN · MovieLens1M ICS | MAE: 0.8411 -> 0.7671 | Helped |
| Add feature completion and three-channel structure completion to LightGCN | FS-GNN · Yelp UCS | MAE: 0.8609 -> 0.8492 | Helped |
| Add feature completion and three-channel structure completion to LightGCN | FS-GNN · Yelp ICS | MAE: 0.8693 -> 0.8002 | Helped |
| Add feature completion modeling (ablation) | FS-GNN · MovieLens100K UCS; MovieLens1M UCS; Yelp UCS; MovieLens100K ICS; MovieLens1M ICS; Yelp ICS | MAE: 0.8359 -> 0.7735; 0.8427 -> 0.7653; 0.9305 -> 0.8492; 0.8805 -> 0.8064; 0.8711 -> 0.7671; 0.8967 -> 0.8002 | Helped |
| Add three-channel structure completion modeling (ablation) | FS-GNN · MovieLens100K UCS; MovieLens1M UCS; Yelp UCS; MovieLens100K ICS; MovieLens1M ICS; Yelp ICS | MAE: 0.7895 -> 0.7735; 0.7887 -> 0.7653; 0.8531 -> 0.8492; 0.8804 -> 0.8064; 0.8731 -> 0.7671; 0.8893 -> 0.8002 | Helped |
Surprising findings
- In the warm-start scenario FS-GNN does not beat all baselines, trailing AGNN and ColdGPT.
Limitations
- Warm-start performance is sub-optimal relative to AGNN and ColdGPT.
- User feature completion relies on up to 30 historical interactions due to token limits.
- Relies on proprietary LLM APIs (gpt-3.5-turbo, ada-002) for feature generation.
Experimental setup
Datasets: MovieLens100K MovieLens1M Yelp
Benchmarks: MovieLens100K MovieLens1M Yelp
Metrics: MAE RMSE
Baselines: NFM DropoutNet LightGCN MetaEmb MeLU MetaHIN MeGNN ColdGPT AGNN
Builds on: LightGCN LLMRec Graph Attention Auto-Encoders AGNN Personalized PageRank
Closest papers by meaning
- Temporally and Distributionally Robust Optimization for Cold-Start Recommendation (AAAI 2024 · similarity 0.86)
- ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.85)
- Graph-enhanced Optimizers for Structure-aware Recommendation Embedding Evolution (NeurIPS 2024 · similarity 0.84)
- SHARP-Distill: A 68× Faster Recommender System with Hypergraph Neural Networks and Language Models (ICML 2025 · similarity 0.84)
- LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.84)
- Preference Aware Dual Contrastive Learning for Item Cold-Start Recommendation (AAAI 2024 · similarity 0.84)
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.