🤖 machine learning

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

DisRFM is a geometry-aware graph domain adaptation framework that leverages polar Riemannian flow matching on constant-curvature manifolds to preserve topological structures and stabilize optimization, thereby outperforming existing Euclidean adversarial methods under semantic and topological domain shifts.

Original authors: Yingxu Wang, Xinwang Liu, Mengzhu Wang, Siyang Gao, Nan Yin

Published 2026-05-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yingxu Wang, Xinwang Liu, Mengzhu Wang, Siyang Gao, Nan Yin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Moving a Team to a New City

Imagine you have a team of expert detectives (a computer program) trained to solve crimes in City A (the Source Domain). They are great at it. Now, you need to send them to City B (the Target Domain) to solve similar crimes.

However, City B is different. The buildings are arranged differently, the streets are wider or narrower, and the layout of the neighborhoods has changed. If you just tell the detectives to "ignore the differences and act like you're in City A," they might get confused. They might forget how to navigate the new street layouts, or they might get so confused by the differences that they stop recognizing the criminals entirely.

This is the problem of Graph Domain Adaptation. The "graphs" are the maps of these cities (nodes are intersections, edges are streets). The goal is to teach the detectives to work in City B without forgetting how to solve crimes.

The Problem: Why Old Methods Fail

Previous methods tried to force the detectives to look exactly the same in both cities. The paper argues this causes two big problems:

Structural Degeneration (The "Blurry Map" Problem):
- The Analogy: Imagine trying to describe a city by only saying, "It's a city." You lose the details. Old methods tried to make the map of City A look exactly like City B by squishing them together. In doing so, they accidentally erased the unique "shape" of the neighborhoods.
- The Result: The detectives lose the ability to tell apart different types of neighborhoods because the "shape" information got washed out.
Optimization Instability (The "Wobbly Tightrope" Problem):
- The Analogy: Training these old models is like trying to balance on a wobbly tightrope while juggling. The computer tries to play a game where one part tries to hide the differences between cities, and another part tries to find them. When the cities are very different, this game becomes chaotic, and the training oscillates (wobbles) without ever finding a stable solution.

The Solution: DisRFM (The "Polar Compass")

The authors propose a new method called DisRFM. Instead of squishing the maps together, they use a special kind of geometry (Riemannian geometry) that treats the maps like a globe rather than a flat piece of paper.

They use a Polar Coordinate System (like a compass with a radius and a direction) to separate the information into two distinct parts:

1. The Radius (The "Size" Dial)

What it is: How far a point is from the center.
The Analogy: Think of this as the size of the neighborhood. Is it a small village or a massive metropolis?
What DisRFM does: It adjusts the "size dial" of the target city to match the source city. If the target city's neighborhoods are generally smaller, DisRFM shrinks the representation so they fit the scale of the source city. This fixes the "Structural Degeneration" by preserving the shape information without erasing it.

2. The Angle (The "Direction" Dial)

What it is: The direction you are facing from the center.
The Analogy: Think of this as the type of neighborhood (e.g., "Industrial," "Residential," "Commercial").
What DisRFM does: It aligns the directions. It makes sure that "Industrial" areas in City A point in the same direction as "Industrial" areas in City B. Crucially, it only does this for areas it is confident about (filtering out uncertain guesses), preventing the detectives from getting confused by bad data.

The Secret Sauce: Flow Matching (The "River Current")

Once the maps are separated into "Size" and "Direction," how do you move the detectives from City A to City B?

Old methods tried to push them directly, which is like throwing a ball through a storm. It's unstable.

DisRFM uses Flow Matching.

The Analogy: Imagine a river flowing smoothly from City A to City B. Instead of throwing the detectives, you place them on a boat that follows the current.
How it works: The system learns the "current" (the vector field) that naturally moves a graph from the source to the target. It ensures the detectives travel along the most efficient, smooth path (a geodesic) that respects the curvature of the globe. This makes the training stable and prevents the "wobbly tightrope" effect.

Why It Works (The Results)

The paper tested this on various datasets (like chemical molecules and protein structures).

The Outcome: DisRFM consistently outperformed all other methods.
The Reason: By separating "Size" (Radius) from "Type" (Angle) and moving them along a smooth, curved path, the system preserved the important structural details while successfully adapting to the new environment.

Summary

Old Way: Squish two different maps together until they look the same, accidentally blurring the details and causing the training to wobble.
DisRFM Way: Put the maps on a globe. Adjust the size (Radius) to match the scale, align the direction (Angle) to match the types, and sail them over using a smooth river current (Flow Matching).

This approach keeps the important structural information intact while successfully transferring knowledge from one domain to another.

Technical Summary: DisRFM

Problem Statement

Graph Domain Adaptation (GDA) aims to transfer graph classifiers from a labeled source domain to an unlabeled target domain under distribution shifts. Unlike standard feature-domain adaptation, GDA faces simultaneous semantic and topological shifts, where graph size, density, degree profiles, and structural roles may vary across domains while the decision boundary must remain transferable.

Existing Euclidean adversarial methods face two primary challenges in this setting:

Structural Degeneration: Euclidean coordinates conflate radial scale and direction. When domain confusion is enforced via adversarial objectives, topology-sensitive variations (which may be entangled with domain variables) are often suppressed alongside domain variation, eroding structural information useful for target prediction.
Optimization Instability: Minimax training between encoders and domain discriminators often induces oscillatory or poorly conditioned gradients, particularly under large structural shifts, leading to unstable training dynamics.

The paper argues that unconditional domain confusion is a structural failure mode; when label-relevant structure correlates with the domain, forcing representations to be indistinguishable can render them uninformative about the label.

Methodology: DisRFM

The authors propose DisRFM (Disentangled Riemannian Flow Matching), a geometry-aware framework that operates on constant-curvature manifolds (e.g., Hyperbolic, Spherical, or Euclidean spaces). Instead of adversarial domain confusion, DisRFM employs Riemannian representation learning and flow-based transport to align domains while preserving label-relevant structure.

The framework consists of two core components:

1. Polar Endpoint Regularization

DisRFM embeds graph representations on a constant-curvature manifold and expresses them in geodesic polar coordinates $(r, u)$ , where $r$ is the geodesic radius and $u$ is the normalized angular direction. This geometry provides an operational bias separating scale-sensitive and direction-sensitive variations:

Radial Structural Calibration: The radius $r$ acts as a proxy for topology-related scale (e.g., graph size/density). DisRFM calibrates the target radial profile to the source by minimizing the univariate Wasserstein distance between sorted radial statistics. This preserves topology-related diversity and prevents structural collapse without forcing all structure into the radius.
Angular Semantic Alignment: The angle $u$ serves as a scale-normalized signal for semantic alignment. Since target labels are unavailable, the method uses source-generated pseudo-labels but filters uncertain predictions via a confidence threshold. The angular loss aligns class prototypes on scale-normalized directions, preserving semantic margins while using the radial magnitude as a soft reliability prior for pseudo-label supervision.

2. Topology-Conditioned Polar Flow Matching

To replace unstable adversarial alignment, DisRFM learns a vector field that transports source embeddings toward target support along class-conditioned geodesics.

Stopped Polar Coupling: For each class, source and reliable target samples are coupled based on a normalized polar transport cost. This cost integrates geometric distance, radial scale difference, angular direction similarity, and target confidence. An entropic transport plan (Sinkhorn) is computed to determine the coupling, with gradients stopped during the coupling step to ensure stability.
Metric-Corrected Flow Regression: A neural network learns a vector field parameterized in polar form (predicting radial speed and angular speed). The flow matching loss minimizes the difference between the learned velocity and the analytic velocity along the geodesic interpolant between coupled pairs. Crucially, the loss incorporates the manifold's warping factor $S_c(r)^2$ to correctly measure angular velocity errors according to the Riemannian metric.

Key Contributions

Identification of Structural Failure Modes: The authors formalize, via information-theoretic arguments, why enforcing unconditional domain invariance can remove label-relevant graph structure when structure and domain are correlated, leading to a lower bound on classification error.
DisRFM Framework: The proposal of a novel framework integrating coordinate-wise radial/angular regularization in Riemannian polar coordinates with geometry-aware flow matching. This approach avoids the pitfalls of Euclidean adversarial training by replacing domain confusion with geometry-constrained, class-conditional manifold alignment.
Theoretical Analysis: The paper provides theoretical guarantees characterizing the structural risk of domain confusion and relating polar discrepancies and flow errors to target risk. It establishes the first-order orthogonality of radial and angular endpoint losses under the polar metric and provides bounds on target risk involving radial, angular, and topology-conditioned transport discrepancies.

Experimental Results

Extensive experiments were conducted on benchmark datasets (Mutagenicity, PROTEINS, NCI1, ogbg-molhiv, DD, BZR, COX2) under diverse domain shifts, including node density, edge density, and feature shifts.

Performance: DisRFM consistently outperforms state-of-the-art baselines, including Euclidean GNNs, manifold-based GNNs, and various GDA methods (adversarial, optimal transport, and manifold-based). The hyperbolic variant (DisRFM-Hn) achieved the best mean performance in 8 out of 12 scenarios, highlighting its advantage in hierarchy-rich structural shifts.
Stability: The method exhibits strong optimization stability, avoiding the gradient oscillations observed in adversarial baselines.
Ablation Studies: Removing flow matching, radial alignment, angular alignment, or the polar disentanglement strategy all resulted in performance degradation, confirming the necessity of each component.
Disentanglement Analysis: Empirical probes demonstrated that the learned radial coordinate primarily correlates with graph size statistics (nodes/edges), while the angular coordinate carries the majority of class-discriminative information, supporting the operational structure–semantics separation.

Significance and Claims

The paper claims that DisRFM addresses the fundamental limitations of existing GDA methods by recognizing that graph domain shifts often entangle label-relevant structural variation with domain-specific nuisance factors. By shifting from Euclidean adversarial alignment to topology-conditioned polar transport, DisRFM successfully aligns domains without erasing structure-sensitive information.

The authors emphasize that their approach does not assume a globally identifiable topology–semantics factorization but rather provides a useful regularization bias for graph transfer. Theoretical analysis supports the design, showing that target risk is controlled by radial discrepancies, angular class-conditional discrepancies, and flow endpoint errors. The work demonstrates that preserving label-relevant structure while transporting samples across semantically and topologically compatible regions is a more robust strategy than unconditional domain confusion.