Research map / Graph learning / Graph self-supervised learning

Empowering Dual-Level Graph Self-Supervised Pretraining with Motif Discovery

AAAI 2024 · Graph learning › Graph self-supervised learning · approach: Semi-supervised node representation learning · 11 citations (OpenAlex)

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 DGPM, a dual-level graph self-supervised pretraining framework that combines GraphMAE-style node feature reconstruction with EdgePool-based motif auto-discovery aligned to WWL graph-kernel similarity, linked by a node-motif cross-matching task.

Problem

Graph self-supervised pretraining suffers from limited topology learning, dependence on hand-defined motifs, and weak interaction between node-level and subgraph-level learning.

Approach

Autonomously discover motifs with EdgePool layers, train motif similarity to match WWL graph-kernel similarity, and add a cross-level node-motif matching discriminator so node and motif encoders interact during pretraining, without predefined motifs.

Key results

• Unsupervised graph classification: DGPM reaches 91.20% on MUTAG, 88.17% on REDDIT-B and 80.87% on NCI1, outperforming listed self-supervised baselines on all 7 datasets.
• Transfer learning on 8 molecular property benchmarks after ZINC15 pretraining: DGPM averages 73.17 ROC-AUC vs 72.64 for GraphMAE, best on 6 of 8 tasks per the authors.
• Ablation: removing motif discovery drops MUTAG from 91.20 to 88.19 and BBBP from 71.2 to 67.4; removing cross-matching costs 0.4-1.5 points.

Measured effects reported in the paper

Surprising findings

• DGPM with automatically discovered motifs matches or slightly beats GROVER and MGSSL, which use predefined chemistry motifs, on average (73.17 vs 72.30 and 73.56 for MGSSL; MGSSL is higher on average, so this is not a clear win over all predefined methods).

Limitations

• Baseline numbers are self-reported from prior papers rather than re-run under identical conditions.
• Motif size depends on the number of EdgePool layers, and the optimal depth varies by dataset.
• Motif discovery benefit is evaluated mainly on small molecular and TU graphs; heterogeneous graphs are left to future work.

Experimental setup

Datasets: ZINC15 IMDB-BINARY IMDB-MULTI PROTEINS COLLAB MUTAG REDDIT-BINARY NCI1 BBBP Tox21 ToxCast SIDER

Benchmarks: IMDB-B IMDB-M PROTEINS COLLAB MUTAG REDDIT-B NCI1 BBBP Tox21 ToxCast SIDER ClinTox

Metrics: accuracy ROC-AUC

Baselines: GIN DiffPool WL DGK graph2vec InfoGraph GraphCL JOAO GCC MVGRL InfoGCL GraphMAE

Builds on: GraphMAE EdgePool Wasserstein Weisfeiler-Lehman graph kernel GraphCL pretrain-gnns

Closest papers by meaning

• Motif-based Graph Self-Supervised Learning for Molecular Property Prediction (NeurIPS 2021 · similarity 0.88)
• GraphGPT: Generative Pre-trained Graph Eulerian Transformer (ICML 2025 · similarity 0.86)
• Beyond Message Passing: Neural Graph Pattern Machine (ICML 2025 · similarity 0.86)
• Equivalence is All: A Unified View for Self-supervised Graph Learning (ICML 2025 · similarity 0.85)
• Exploring Correlations of Self-Supervised Tasks for Graphs (ICML 2024 · similarity 0.85)
• HGPrompt: Bridging Homogeneous and Heterogeneous Graphs for Few-Shot Prompt Learning (AAAI 2024 · similarity 0.85)