Research map / Graph learning / Graph self-supervised learning
Sterling: Synergistic Representation Learning on Bipartite Graphs
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 Sterling, a non-contrastive, bootstrapped SSL model for bipartite graphs that combines inter-/intra-type local positive pairs with a mutual-information co-clustering objective, requiring no negative pairs.
Problem
Learning informative node embeddings for bipartite graphs; existing bipartite SSL methods rely on contrastive learning with many negative pairs, which is computationally costly and can produce semantic errors.
Approach
A BYOL-style online/target encoder learns embeddings by maximizing similarity of inter-type (connected user-item) and intra-type (k-NN, Adamic-Adar weighted) positive pairs, plus a global loss that maximizes mutual information between learned co-cluster distributions; a learned joint distribution p(U,V) combines metapath adjacency and embedding-derived adjacency. Theory claims I(K;L) lower-bounds
Key results
- On ML-100K top-10 recommendation, Sterling reaches F1@10 25.54 vs. 25.19 for the strongest baseline SimGCL.
- On Wikipedia (40%) link prediction, Sterling achieves AUC 95.04 vs. 94.62 for SimGCL.
- On co-clustering, Sterling improves NMI on IMDB to 33.4 vs. 26.8 for DeepCC, and ACC on Cornell to 73.4 vs. 68.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add global objective L_glb (ablation) | Sterling · Cornell | NMI: 26.02 -> 37.51 | Helped |
| Add inter-type objective L_uv (ablation) | Sterling · ML-100K | F1@10: 0.20 -> 25.54 | Helped |
| Add intra-type objective L_v (ablation) | Sterling · ML-100K | F1@10: 21.67 -> 25.54 | Helped |
| Add intra-type objective L_u (ablation) | Sterling · Cornell | NMI: 35.74 -> 37.51 | Helped |
| Add metapath adjacency A_meta (ablation) | Sterling · Cornell | NMI: 26.61 -> 37.51 | Helped |
| Add semantic adjacency A_emb (ablation) | Sterling · Wikipedia (40%) | AUC: 94.42 -> 95.04 | Helped |
| Use noise filtering (ablation) | Sterling · Cornell | NMI: 36.82 -> 37.51 | Helped |
| Add the Adamic-Adar index s_aa (ablation) | Sterling · ML-100K | F1@10: 20.32 -> 25.54 | Helped |
| Add semantic similarity s_emb (ablation) | Sterling · ML-100K | F1@10: 24.84 -> 25.54 | Helped |
| Replace absolute activation with ReLU (ablation) | Sterling · Cornell | NMI: 37.51 -> 36.53 | Hurt |
| Replace Adamic-Adar with Co-HITS (ablation) | Sterling · Cornell | NMI: 37.51 -> 35.87 | Hurt |
| Use only U_theta V_phi^T for A_emb (ablation) | Sterling · Cornell | NMI: 37.51 -> 36.92 | Hurt |
Surprising findings
- Removing the inter-type loss L_uv collapses ML-100K F1@10 to 0.20, while intra-type and global losses matter more for the simpler link prediction and co-clustering tasks.
- Using absolute activation for the joint distribution adjacency outperforms ReLU, which the authors attribute to datasets not distinguishing sign of connections.
Limitations
- Hyperparameters (k-NN, number of co-clusters, threshold alpha) are selected by grid search and sensitivity analysis, with sensitivity reported only on Cornell.
- Theoretical result (I(K;L) lower bound on I(U;V)) relies on assumptions such as neural networks being deterministic and injective.
- Evaluated only on relatively small bipartite datasets (at most ~15,000 x 3,214 nodes).
Experimental setup
Datasets: ML-100K Wikipedia IMDB Cornell Citeseer
Benchmarks: ML-100K Wikipedia IMDB Cornell Citeseer
Metrics: F1@10 NDCG MAP MRR AUC NMI ACC
Baselines: SimGCL COIN BiGI DirectAU NGCF NeuMF AFGRL GraphMAE DeepCC HDMI DMGI HeCo
Builds on: BYOL BGRL DeepCC SimGCL Adamic-Adar index
Closest papers by meaning
- Equivalence is All: A Unified View for Self-supervised Graph Learning (ICML 2025 · similarity 0.87)
- S3GCL: Spectral, Swift, Spatial Graph Contrastive Learning (ICML 2024 · similarity 0.86)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.85)
- Learning Representations for Hierarchies with Minimal Support (NeurIPS 2024 · similarity 0.85)
- Beyond Homophily: Graph Contrastive Learning with Macro-Micro Message Passing (AAAI 2025 · similarity 0.85)
- Transitivity-Preserving Graph Representation Learning for Bridging Local Connectivity and Role-Based Similarity (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.