Research map / Label-efficient and robust learning / Recognition and long-tail
ContextGNN: Beyond Two-Tower Recommendation Systems
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 ContextGNN, a hybrid GNN recommender that combines NBFNet-style pair-wise representations for items in a user's local subgraph with a shallow-item-embedding two-tower model for distant items, fused by a learned user-specific score.
Problem
Two-tower recommenders learn pair-agnostic user and item representations, while pair-wise models scale quadratically or are restricted to candidates within the user's local subgraph, limiting coverage of exploratory items.
Approach
A single GNN backbone produces both pair-wise scores for items inside the user's k-hop subgraph (NBFNet-style readout) and two-tower scores via shallow item embeddings for items outside it; a user-specific MLP fusion score merges them into one ranking trained end-to-end.
Key results
- On RelBench, ContextGNN averages 9.23 test MAP versus 7.71 for NBFNet (best pair-wise baseline), about 20% improvement, and far above the best two-tower baseline (about 344% improvement).
- On site-sponsor-run (rel-trial), ContextGNN roughly doubles performance over the individual components (test 28.02 vs 19.06 NBFNet).
- On the IJCAI Contest dataset, ContextGNN reaches HR@1 0.411 versus 0.148 for the best baseline TGT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse pair-wise local scores with two-tower scores for distant items | ContextGNN vs. NBFNet · RelBench recommendation tasks | MAP: 20% on average | Helped |
| Fuse pair-wise local scores with two-tower scores for distant items | ContextGNN vs. best two-tower baseline · RelBench recommendation tasks | MAP: 344% on average | Helped |
| Fuse pair-wise and two-tower recommendation signals | ContextGNN vs. individual components · rel-trial site-sponsor-run | MAP: ≈ 100% | Helped |
| Use ContextGNN instead of two-tower recommendation models | ContextGNN · IJCAI Contest | HR@1: ~170% improvement | Helped |
| Use ContextGNN instead of the strongest specialized baseline | ContextGNN · Amazon-Book | Recall@20: 0.0681 -> 0.0451 | Hurt |
| Use ContextGNN instead of the strongest specialized baseline | ContextGNN · Amazon-Book | NDCG@20: 0.0556 -> 0.0377 | Hurt |
| Use ContextGNN instead of two-tower baselines | ContextGNN · RelBench recommendation tasks | MAP: ContextGNN test average: 9.23; best two-tower test average: 2.08 | Helped |
| Increase subgraph depth from k=1 to k=3 | locality score · RelBench recommendation tasks | locality score | Helped |
Surprising findings
- On static Amazon-Book, ContextGNN (Recall@20 0.0451) underperforms UltraGCN (0.0681) and LightGCL (0.0585), though it beats NGCF and LightGCN.
- Shallow item embeddings in the two-tower ShallowItem model underperform GraphSAGE and NGCF on the condition-sponsor-run task, suggesting deep item-side GNNs matter there.
Limitations
- Transductive shallow item embeddings cannot handle new items at prediction time without the proposed inductive extension.
- Underperforms some simplified or self-supervised static GNN recommenders on Amazon-Book.
- Locality computation and subgraph sampling are limited to depth k=3 for scalability.
Experimental setup
Datasets: RelBench rel-amazon rel-hm rel-stack rel-trial Amazon-Book IJCAI Contest
Benchmarks: RelBench Amazon-Book IJCAI Contest
Metrics: MAP@k Recall@20 NDCG@20 HitRate@k NDCG@k runtime
Baselines: LightGBM MultiVAE GraphSAGE NGCF NBFNet ShallowItem UltraGCN LightGCL SimGCL SGL DeepFM Bert4Rec
Builds on: NBFNet GraphSAGE LightGCN Relational Deep Learning
Closest papers by meaning
- LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.88)
- SHARP-Distill: A 68× Faster Recommender System with Hypergraph Neural Networks and Language Models (ICML 2025 · similarity 0.86)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.86)
- Semantic Enhanced Heterogeneous Hypergraph Network for Collaborative Filtering (AAAI 2025 · similarity 0.86)
- Spectral-Based Graph Neural Networks for Complementary Item Recommendation (AAAI 2024 · similarity 0.86)
- Beyond Graph Convolution: Multimodal Recommendation with Topology-aware MLPs (AAAI 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.