Research map / Graph learning / GNN architectures
Improving Distinguishability of Class for Graph Neural Networks
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 Disc-GNN, a GNN whose layers are guided by local and global "distinguishability of class" metrics (LDC/GDC) via LDC-based inter-layer filtering, initial compensation, and a GDC regularizer, improving deep node classification and clustering.
Problem
Performance of GNNs degrades as layers are stacked (over-smoothing and heterophily); the paper seeks to keep node representations distinguishable across classes in deep GNNs.
Approach
Defines LDC (per-node gap between max and min class probabilities) and GDC (its graph average) as indirect measures of representation quality, then uses LDC to gate inter-layer representations, adds LDC-adaptive initial residual compensation, and adds a GDC regularization term to the training objective.
Key results
- Disc-GNN reaches 88.73% ACC on Cora and 85.45% on Wisconsin for node classification, the best among the compared methods in Table 2.
- On node clustering with k-means, Disc-GNN improves NMI by 4.93% and ARI by 4.92% over GCN on Cora.
- Pearson correlation between GDC and ACC across GCN depths is reported as 0.9968, higher than for global smoothness (0.7877).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Disc-GNN instead of GCNII | Node classification models · Cora | ACC (%): 88.09 ±1.56 -> 88.73 ±1.48 | Helped |
| Use Disc-GNN instead of GCNII | Node classification models · Citeseer | ACC (%): 75.97 ±1.86 -> 76.65 ±1.11 | Helped |
| Use Disc-GNN instead of JK-Net | Node classification models · Pubmed | ACC (%): 88.60 ±0.59 -> 88.84 ±0.55 | Helped |
| Use Disc-GNN instead of DeCorr | Node classification models · Wisconsin | ACC (%): 82.23 ±6.15 -> 85.45 ±2.51 | Helped |
| Use Disc-GNN instead of DeCorr | Node classification models · Texas | ACC (%): 79.00 ±6.02 -> 81.45 ±4.59 | Helped |
| Use Disc-GNN instead of DeCorr | Node classification models · Cornell | ACC (%): 81.43 ±6.65 -> 83.89 ±4.26 | Helped |
| Use Disc-GNN instead of APPNP | Node clustering models · Cora | NMI (%): 78.89 ±0.84 -> 80.66 ±1.08 | Helped |
| Use Disc-GNN instead of GCNII | Node clustering models · Texas | NMI (%): 65.49 ±3.45 -> 69.43 ±2.07 | Helped |
| Use Disc-GNN instead of GCNII | Node clustering models · Wisconsin | NMI (%): 46.96 ±3.39 -> 70.87 ±2.34 | Helped |
| Add initial compensation based on LDC (ablation) | Disc-GNN · Cora and Texas, 10 and 20 layers | node classification accuracy | Helped |
| Add global optimization based on GDC (ablation) | Disc-GNN · Cora and Texas, 10 and 20 layers | node classification accuracy | Helped |
| Add inter-layer filtering based on LDC (ablation) | Disc-GNN · Texas | node classification accuracy | Helped |
Surprising findings
- Global smoothness rises sharply at around the 8th GCN layer before dropping, suggesting over-smoothing only appears after a certain depth.
Limitations
- Relies on the class-probability mapping matrix, so LDC/GDC are indirect measures of representation quality and are only defined for classification-style outputs.
- Several hyperparameters (dropout, weight decay, eta, epsilon) are tuned over ranges, and the regularization coefficient and relaxation factor add sensitivity.
Experimental setup
Datasets: Cora Citeseer Pubmed Wisconsin Texas Cornell
Benchmarks: Cora Citeseer Pubmed Wisconsin Texas Cornell
Metrics: ACC NMI ARI
Baselines: GCN GAT SGC JK-Net APPNP DAGNN DeCorr FAGCN GCNII
Builds on: GCN GCNII APPNP DeCorr
Closest papers by meaning
- How Powerful are Graph Neural Networks? (ICLR 2019 · similarity 0.90)
- Exploring Consistency in Graph Representations: from Graph Kernels to Graph Neural Networks (NeurIPS 2024 · similarity 0.88)
- Graph Classification via Reference Distribution Learning: Theory and Practice (NeurIPS 2024 · similarity 0.88)
- Classic GNNs are Strong Baselines: Reassessing GNNs for Node Classification (NeurIPS 2024 · similarity 0.88)
- GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights (NeurIPS 2025 · similarity 0.88)
- Integrating Co-Training with Edge Discrimination to Enhance Graph Neural Networks Under Heterophily (AAAI 2025 · similarity 0.88)
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.