Research map / Label-efficient and robust learning / Recognition and long-tail
Cooperation of Experts: Fusing Heterogeneous Information with Large Margin
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 Cooperation of Experts (CoE), which fuses multiplex/multimodal networks with graph structure learning, two-level experts trained with mutual-information lower bounds, and a large-margin confidence-tensor mechanism for expert collaboration.
Problem
Fusing heterogeneous multi-typed information (multiplex networks and multimodal data) for node classification, where a single predictor across all layers ignores heterogeneous link patterns and existing MoE methods favor expert competition over cooperation.
Approach
Replaces gating-based expert competition with cooperation: low- and high-level experts trained on individual and fused graphs with mutual-information objectives, and a learnable confidence tensor that combines all experts' outputs and is optimized with a large-margin loss between the top-two predicted classes.
Key results
- CoE achieves the highest node classification accuracy on all five multiplex benchmarks, e.g., 94.21% on ACM and 92.27% on DBLP, 93.40% on Yelp, 78.37% on MAG, 98.01% on Amazon.
- On multimodal datasets CoE reports 81.11% (ESP), 70.24% (Flickr), 71.04% (IAPR), 66.80% (NUS), exceeding the best baselines (e.g., QMF 69.08% on IAPR, 65.42% on NUS).
- Replacing the large-margin mechanism with RF or WRF lowers accuracy (e.g., ACM 93.39% RF, 93.64% WRF vs 94.21% CoE); removing high-level experts drops Yelp to 68.27%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CoE instead of the strongest reported baseline | CoE · ACM, DBLP, Yelp, MAG, Amazon | accuracy (%): 94.21±0.14 vs 92.81±0.29 (ACM); 92.27±0.24 vs 91.49±0.32 (DBLP); 93.40±0.12 vs 92.01±0.42 (Yelp); 78.37±0.16 vs 77.40±0. | Helped |
| use CoE instead of the strongest reported baseline | CoE · ESP, Flickr, IAPR, NUS | accuracy (%): 81.11±0.05 vs 80.19±0.56 (ESP); 70.24±0.09 vs 69.49±0.46 (Flickr); 71.04±0.04 vs 69.08±0.45 (IAPR); 66.80±0.32 vs 65.42± | Helped |
| add high-level experts (ablation) | CoE · ACM, DBLP, Yelp, Amazon | accuracy (%): 91.25±0.17 -> 94.21±0.14 (ACM); 90.71±0.26 -> 92.27±0.24 (DBLP); 68.27±0.32 -> 93.40±0.12 (Yelp); 94.50±0.16 -> 98.01±0. | Helped |
| add high-level experts (ablation) | CoE · MAG | accuracy (%): 78.05±0.27 -> 78.37±0.16 | Hurt |
| add graph structure learning (ablation) | CoE · ACM, DBLP, Yelp, MAG | accuracy (%): 93.60±0.29 -> 94.21±0.14 (ACM); 91.13±0.24 -> 92.27±0.24 (DBLP); 93.14±0.33 -> 93.40±0.12 (Yelp); 77.97±0.34 -> 78.37±0. | Helped |
| add graph structure learning (ablation) | CoE · Amazon | accuracy (%): 97.87±0.24 -> 98.01±0.09 | Hurt |
| use large-margin expert collaboration instead of random forest (ablation) | CoE · ACM, DBLP, Yelp, MAG, Amazon | accuracy (%): 93.39±0.19 -> 94.21±0.14 (ACM); 91.48±0.14 -> 92.27±0.24 (DBLP); 91.61±0.21 -> 93.40±0.12 (Yelp); 76.91±0.16 -> 78.37±0. | Helped |
| use large-margin expert collaboration instead of weighted random forest (ablation) | CoE · ACM, DBLP, Yelp, MAG, Amazon | accuracy (%): 93.64±0.15 -> 94.21±0.14 (ACM); 91.97±0.22 -> 92.27±0.24 (DBLP); 93.05±0.16 -> 93.40±0.12 (Yelp); 77.45±0.18 -> 78.37±0. | Helped |
| use weighted random forest instead of random forest (ablation) | CoE expert collaboration · ACM, DBLP, Yelp, MAG, Amazon | accuracy (%): 93.39±0.19 -> 93.64±0.15 (ACM); 91.48±0.14 -> 91.97±0.22 (DBLP); 91.61±0.21 -> 93.05±0.16 (Yelp); 76.91±0.16 -> 77.45±0. | Helped |
Surprising findings
- Removing high-level experts causes a much larger drop than removing graph structure learning, e.g., Yelp falls from 93.40% to 68.27%.
- Several baselines (e.g., Mowst on ACM at 85.69%) score below GCN on some datasets, suggesting instability of existing expert-based methods.
Limitations
- Theoretical guarantees (Lipschitz constant, convergence) rely on assumptions and bounds, and the convexity claim is only partial with respect to intermediate variables.
- Baseline comparisons use GCN as the backbone and report maximum accuracy over single networks for methods not supporting multiplex data, which may limit fairness.
- Several baselines run out of memory on large datasets (MAG, Amazon), limiting comparison.
Experimental setup
Datasets: ACM DBLP Yelp MAG Amazon ESP Flickr IAPR NUS
Benchmarks: ACM DBLP Yelp MAG Amazon ESP Flickr IAPR NUS
Metrics: accuracy standard deviation
Baselines: GCN HAN LDS GRCN IDGL ProGNN GEN NodeFormer SUBLIME STABLE GSR HDMI
Builds on: Graph Structure Learning (GSL) Mixture of Experts contrastive mutual information estimation (InfoNCE) GCN
Closest papers by meaning
- Unifying Multi-View Knowledge for Graph Learning via Model Collaboration (AAAI 2026 · similarity 0.86)
- Atomic HINs: Entity-Attribute Duality for Heterogeneous Graph Modeling (ICLR 2026 · similarity 0.85)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.85)
- Learning to Specialize: Joint Gating-Expert Training for Adaptive MoEs in Decentralized Settings (NeurIPS 2025 · similarity 0.84)
- Harnessing Language Model for Cross-Heterogeneity Graph Knowledge Transfer (AAAI 2025 · similarity 0.84)
- Heterogeneous Graph Neural Network on Semantic Tree (AAAI 2025 · 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.