Research map / Graph learning / Graph self-supervised learning
Rethinking Dimensional Rationale in Graph Contrastive Learning from Causal Perspective
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 DRGCL, a graph contrastive learning method that learns a dimensional rationale (a learnable per-dimension weight on graph embeddings) via bi-level meta-learning plus a redundancy-reduction loss, motivated by an SCM and backdoor adjustment.
Problem
Graph contrastive learning captures invariant information but may learn noisy, task-agnostic information; prior rationale methods focus on structural rationales (SR), which can hurt discriminability.
Approach
Instead of structural rationales, learn a dimensional rationale (a learnable weight over embedding dimensions) updated by bi-level meta-learning so that it acts as a backdoor-adjusted confounder, combined with a decorrelation (redundancy reduction) loss on the embeddings.
Key results
- On TU unsupervised benchmarks DRGCL reaches an average rank of 2.8, the best among compared methods (RGCL and SimGRACE 3.3, GraphCL 5.0).
- On ZINC-2M transfer to eight MoleculeNet datasets DRGCL achieves 72.5 average ROC-AUC, versus 71.8 for RGCL and 72.1 for JOAOv2.
- Ablation on transfer learning: removing the DR module lowers the average from 72.5 to 71.9, and removing the redundancy reduction lowers it to 71.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add learnable dimensional rationale and redundancy reduction | GraphCL · ZINC-2M transfer benchmarks | average ROC-AUC: 70.8 -> 72.5 | Helped |
| add learnable dimensional rationale (ablation) | DRGCL · ZINC-2M transfer benchmarks | average ROC-AUC: 71.9 -> 72.5 | Helped |
| add redundancy reduction (ablation) | DRGCL · ZINC-2M transfer benchmarks | average ROC-AUC: 71.0 -> 72.5 | Helped |
| add dimensional rationale to GraphCL (ablation) | GraphCL · ZINC-2M transfer benchmarks | average ROC-AUC: 70.8 -> 71.9 | Helped |
| add redundancy reduction to GraphCL (ablation) | GraphCL · ZINC-2M transfer benchmarks | average ROC-AUC: 70.8 -> 71.0 | Helped |
| replace adaptive dimensional weights with fixed R=0.1 (ablation) | DRGCL · ZINC-2M transfer benchmarks | average ROC-AUC: 71.2 -> 70.5 | Hurt |
| replace adaptive dimensional weights with fixed R=0.3 (ablation) | DRGCL · ZINC-2M transfer benchmarks | average ROC-AUC: 71.2 -> 71.0 | Hurt |
| replace adaptive dimensional weights with fixed R=0.7 (ablation) | DRGCL · ZINC-2M transfer benchmarks | average ROC-AUC: 71.2 -> 70.2 | Hurt |
| replace adaptive dimensional weights with fixed R=1.0 (ablation) | DRGCL · ZINC-2M transfer benchmarks | average ROC-AUC: 71.2 -> 70.5 | Hurt |
| increase training epochs (ablation) | DRGCL · REDDIT-BINARY | accuracy | Helped |
| increase batch size up to 128 (ablation) | DRGCL · NCI1 and REDDIT-BINARY | accuracy | Helped |
| increase batch size beyond 128 (ablation) | DRGCL · NCI1 and REDDIT-BINARY | accuracy | Hurt |
Surprising findings
- Fixed dimensional weights (0.1 to 1.0) also outperform GraphCL-style baselines on transfer, and the authors find that random dimension preservation can improve or degrade performance, which they interpret as the DR acting as a confounder.
Limitations
- Bi-level meta-learning makes training slower than GraphCL.
- The method is described as plug-and-play but only evaluated with GIN and graph contrastive pretraining; the theoretical guarantees rely on idealized assumptions (optimal R, best decorrelation).
Experimental setup
Datasets: TU datasets (NCI1, PROTEINS, DD, MUTAG, COLLAB, RDT-B, RDT-M5K, IMDB-B) ZINC-2M PPI-306K MoleculeNet (BBBP, Tox21, ToxCast, SIDER, ClinTox, MUV, HIV, BACE)
Benchmarks: TU datasets MoleculeNet PPI-306K
Metrics: classification accuracy ROC-AUC average rank
Baselines: GraphCL ADGCL JOAO JOAOv2 RGCL SimGRACE InfoGraph AttrMasking ContextPred graph2vec
Builds on: GraphCL RGCL SimGRACE InfoNCE Barlow Twins
Closest papers by meaning
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.91)
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.88)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.87)
- A New Mechanism for Eliminating Implicit Conflict in Graph Contrastive Learning (AAAI 2024 · similarity 0.87)
- Scores for Learning Discrete Causal Graphs with Unobserved Confounders (AAAI 2024 · similarity 0.87)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · 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.