Research map / Graph learning / Graph self-supervised learning

DisRFM: Polar Riemannian Flow Matching for Structure-Preserving Graph Domain Adaptation

NeurIPS 2026 · Graph learning › Graph self-supervised learning · approach: Graph domain adaptation · 0 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 DisRFM, a geometry-aware graph domain adaptation method that uses polar endpoint regularization (radial Wasserstein, confidence-filtered angular alignment) and topology-conditioned Riemannian flow matching on constant-curvature manifolds instead of adversarial domain confusion.

Problem

Graph domain adaptation for graph classification under combined semantic and topological shifts, where adversarial Euclidean alignment erodes structure information and yields unstable minimax training.

Approach

Embed graphs on a constant-curvature manifold, express them in geodesic polar coordinates, align radii via univariate Wasserstein and angles via confidence-filtered class-conditional alignment, and transport source to target with a polar-metric-corrected flow matching loss over class-compatible Sinkhorn couplings.

Key results

• On Mutagenicity node-shift M0->M1, DisRFM-H^n reaches 79.8% vs. 79.3% for GAA (best baseline in that column).
• On feature shift P->D, DisRFM-S^n reaches 63.8% vs. 62.4% for GAA.
• Ablation on Mutagenicity: full DisRFM reaches 78.3 on M0->M1 vs. 75.9 for w/o PE, 76.2 for w/o FM.

Surprising findings

• Table shows the hyperbolic variant wins only in 8 of 12 scenarios, with spherical and Euclidean variants competitive, suggesting gains are not solely from curvature choice.

Limitations

• Theoretical guarantees are conditional and do not establish identifiable topology-semantics separation or global convergence.
• Minibatch Sinkhorn coupling and pseudo-label errors are approximation terms not fully controlled.
• Gains over strongest manifold-based GDA baselines are often small (about 0.5 points) in the reported tables.

Experimental setup

Datasets: Mutagenicity PROTEINS NCI1 ogbg-molhiv DD BZR BZR_MD COX2 COX2_MD

Benchmarks: Mutagenicity DD PROTEINS BZR BZR_MD COX2 COX2_MD

Metrics: classification accuracy (%)

Baselines: WL subtree PathNN GCN GIN CIN GMT dDGM RieGrace ProGDM D-GCN DEAL SGDA

Builds on: hyperbolic graph neural networks riemannian flow matching adversarial domain adaptation sinkhorn optimal transport

Closest papers by meaning

• Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.88)
• TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.87)
• Pairwise Alignment Improves Graph Domain Adaptation (ICML 2024 · similarity 0.87)
• Disentangled Graph Spectral Domain Adaptation (ICML 2025 · similarity 0.87)
• Flowette: Flow Matching with Graphette Priors for Graph Generation (NeurIPS 2026 · similarity 0.87)
• Toward a Unified Geometry Understanding : Riemannian Diffusion Framework for Graph Generation and Prediction (NeurIPS 2025 · similarity 0.86)