Research map / Graph learning / Graph self-supervised learning

Every Node Is Different: Dynamically Fusing Self-Supervised Tasks for Attributed Graph Clustering

AAAI 2024 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs · 15 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 DyFSS, which fuses embeddings from five self-supervised tasks per node via a mixture-of-experts gating network, trained with dual-level pseudo-label and graph-structure supervision, for attributed graph clustering.

Problem

Existing multi-task SSL methods for attributed graph clustering share one set of SSL task weights across all nodes, which ignores that different nodes need different emphasis on tasks.

Approach

Use a mixture-of-experts network where each expert is a GCN layer for one SSL task and a per-node gating network produces fusion weights, trained with a pseudo-label loss and an adjacency-reconstruction loss to make fusion reliable early in training.

Key results

• DyFSS achieves ACC 72.19 on Cora, versus 69.57 for AUTOSSL, and is best on all five datasets for ACC in the reported table except CS.
• On Photo, DyFSS reaches 79.81 ACC, improving up to 8.66% over state-of-the-art multi-task SSL methods per the abstract.
• Ablation: removing the dual-level supervision L_dual lowers ACC on Computers from 56.95 to 47.96.

Measured effects reported in the paper

Surprising findings

• Fixed-weight multi-task SSL (AUTOSSL) prefers the topology task and fails to capture the clustering structure on Cora, per the visualization analysis.
• On the CS dataset, DyFSS does not achieve the best ACC (75.81 vs 73.31 for CCA-SSG is runner-up; the table shows DyFSS 75.81 vs best-reported lower bar in other columns), showing gains are not uniform.

Limitations

• Reported gains are not uniform across datasets; e.g., on CS, DyFSS ACC is not the top value in the table.
• Depends on a two-stage pretraining pipeline with five SSL tasks and several hyperparameters (lambda_1, lambda_2, m) tuned per dataset.

Experimental setup

Datasets: Cora Citeseer Photo Computers CS

Benchmarks: Cora Citeseer Photo Computers CS

Metrics: ACC NMI F1

Baselines: K-means AE DeepWalk VGAE ARVGA DAEGC GraphMAE DGI GCA CCA-SSG BGRL AUTOSSL

Builds on: AUTOSSL ARVGA PARETOGNN DEC-style clustering alignment

Closest papers by meaning

• Revisiting Self-Supervised Heterogeneous Graph Learning from Spectral Clustering Perspective (NeurIPS 2024 · similarity 0.89)
• DGCLUSTER: A Neural Framework for Attributed Graph Clustering via Modularity Maximization (AAAI 2024 · similarity 0.88)
• Exploring Correlations of Self-Supervised Tasks for Graphs (ICML 2024 · similarity 0.88)
• Clustering Properties of Self-Supervised Learning (ICML 2025 · similarity 0.87)
• Local Graph Clustering with Noisy Labels (ICLR 2024 · similarity 0.86)
• Federated Graph-Level Clustering Network (AAAI 2025 · similarity 0.86)