Research map / Graph learning / Graph self-supervised learning

Chasing Label Shifters: A Change-Aware Framework for Dynamic Graph Node Classification

NeurIPS 2026 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs

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.

Identifies a shifter gap (nodes whose labels change across snapshots are misclassified, often stuck on the old label) in dynamic GNNs and proposes CHASE, a model-agnostic head-level wrapper with a homophily-guided contrastive loss, change detection head, and change-aware classifier, plus three new dynamic node benchmarks.

Problem

Dynamic GNNs classify nodes whose labels change between snapshots (shifters) much worse than stable nodes, with errors concentrated on the old label.

Approach

Attributes the shifter failure to embedding inertia from neighborhood aggregation and wraps any dynamic GNN with a contrastive loss pulling shifter embeddings toward the new class, a change detection head estimating shift probability, and a classifier that consumes this change signal and the previous prediction to override stale predictions.

Key results

• CHASE improves shifter accuracy and F1 across all 21 backbone-dataset combinations, with up to 114.9% accuracy and 202.5% F1 relative improvement.
• Ablation on GCN-GRU shows the change detection head gives the largest single jump in shifter performance (e.g., Music shifter F1 from 56.7 to 75.3).
• Change detection head reaches AUROC up to 97.4% for shifter detection on GossipCop.

Surprising findings

• A rule-based variant that flips the prediction when pchange > 0.5 degrades sharply on the multiclass Music dataset (shifter F1 18.6 vs 46.8 baseline).
• Even sequential static GCN and EERM benefit substantially from CHASE augmentation, and EERM fails to improve over GCN.

Limitations

• Treats label change as binary, ignoring gradual transitions near the decision boundary.
• Operates only on discrete snapshots; not extended to continuous-time dynamic graphs.
• Relies on the model's own previous predictions (hproxy and y_hat(t-1)), which may be incorrect.

Experimental setup

Datasets: DBLP-S GossipCop Music

Benchmarks: DBLP-S GossipCop Music

Metrics: accuracy macro F1 AUROC

Baselines: GCN EERM GCN-GRU EvolveGCN DySAT GCN-SE IDGNN TADGNN SiGNN

Builds on: GCN-GRU EvolveGCN DySAT GCN-SE IDGNN

Closest papers by meaning

• Dynamic Graph Neural Networks Under Spatio-Temporal Distribution Shift (NeurIPS 2022 · similarity 0.88)
• Towards OOD Generalization in Dynamic Graphs via Causal Invariant Learning (AAAI 2026 · similarity 0.88)
• Adaptive Mixture of Disentangled Experts for Dynamic Graphs under Distribution Shifts (ICLR 2026 · similarity 0.88)
• Matcha: Mitigating Graph Structure Shifts with Test-Time Adaptation (ICLR 2025 · similarity 0.87)
• Topology-Aware Dynamic Reweighting for Distribution Shifts on Graph (ICML 2025 · similarity 0.86)
• Towards Dynamic Message Passing on Graphs (NeurIPS 2024 · similarity 0.86)