🤖 machine learning

Cross-Resolution Semantic Learning for Graph Domain Adaptation

This paper proposes Cross-Resolution Semantic Learning (CReSL), a novel Graph Domain Adaptation method that addresses semantic resolution shifts by learning soft source-to-target resolution correspondences through multi-resolution representation banks, prototype transport, and target grafting to optimize knowledge transfer across varying neighborhood ranges.

Original authors: Yingxu Wang, Haoze Huang, Zhongkai Zheng, Shangsong Liang

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yingxu Wang, Haoze Huang, Zhongkai Zheng, Shangsong Liang

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

Imagine you are trying to teach a student from one country how to play a sport in a completely different country. The rules of the game (the "classes" or categories) are the same, but the field is different: the grass is taller, the wind blows harder, and the ball bounces differently. In the world of artificial intelligence, this is called Graph Domain Adaptation. Think of a "graph" as a map of connections—like a social network where people are dots and friendships are lines. AI models learn to predict things (like whether a person likes a certain movie) by looking at these connections.

The tricky part is that what works on the "home field" (the source data) doesn't always work on the "away field" (the target data). Sometimes, to understand a friend's interests, you need to look at their immediate best friends (a short distance). But on a different field, you might need to look at their friends' friends, or even their friends' friends' friends (a longer distance) to get the same clue. This paper tackles a specific problem: what if the "right distance" to look changes when you switch fields? The authors call this a semantic resolution shift. It's like if your teacher told you to study the neighborhood around your house to find a lost toy, but when you move to a new city, the toy is actually hidden three blocks away. If you keep looking in the same-sized circle, you'll never find it.

The Problem: Looking at the Wrong Distance

The authors noticed that existing AI methods often make a simple mistake. They assume that if you learn something by looking at a "1-hop" neighborhood (immediate neighbors) in the source graph, you should look at the "1-hop" neighborhood in the target graph. But the paper argues that this is often wrong. Just because the meaning of the data is the same doesn't mean the distance needed to understand it is the same.

They call the distance an AI looks at its "propagation resolution." In some cases, the most important clues for a specific category might be found at a 2-hop distance in the source data, but at a 4-hop distance in the target data. If the AI blindly pairs them up (1-hop with 1-hop, 2-hop with 2-hop), it's like trying to fit a square peg in a round hole. The result is "negative transfer," where the AI actually gets worse because it's learning from the wrong perspective.

The Solution: CReSL, the Flexible Detective

To fix this, the researchers proposed a new method called CReSL (Cross-Resolution Semantic Learning). Instead of forcing a rigid match, CReSL acts like a flexible detective that learns to match the right "viewing distance" from the source to the best "viewing distance" in the target.

Here is how it works, broken down into three playful steps:

The Multi-Resolution Bank: Imagine the AI has a set of different camera lenses. One lens zooms in tight (0-hop), one looks a bit further (1-hop), and others zoom out even more (2-hop, 4-hop, etc.). The AI takes a picture of the graph through every lens at once. This creates a "bank" of representations. Crucially, it keeps these views separate so it doesn't blur the details, but it uses a shared brain to understand them all.
The Prototype Transport (The Matchmaker): This is the clever part. The AI looks at the "prototypes" (the average examples) of different categories in the source data. It then asks: "If I look at the target data through the 2-hop lens, does it look like the source data seen through the 4-hop lens?" It doesn't need to know the exact answers (labels) for the target data to figure this out. It uses a soft, probabilistic guess to see which source lens matches which target lens best. It's like saying, "The view from the 2-hop lens in the new city feels most similar to the 4-hop view from the old city, so let's use that connection."
The Target Grafting (The Tailor): Once the AI knows which lenses match, it doesn't just swap them; it "grafts" the knowledge. It takes a specific target graph and gently nudges it toward the source's structure, but only as much as the AI is confident about the category. If the AI is unsure if a graph belongs to "Class A" or "Class B," it mixes the adjustments for both classes. This prevents the AI from making a hard, potentially wrong guess. It's like a tailor adjusting a suit: if you aren't sure if the customer is tall or short, you make a small, balanced adjustment rather than cutting the fabric off entirely.

What They Found

The researchers tested CReSL on various graph datasets, including chemical molecules and protein structures, where they artificially changed the structure or features to simulate moving to a new "domain."

The results suggest that CReSL is quite effective. In their experiments, it consistently outperformed other strong methods. For instance, on a dataset called Mutagenicity, when moving from one structural shift to another (M0→M1), CReSL achieved 83.0% accuracy, beating the next best method (GAA) which got 79.3%. On another shift (M0→M2), it hit 74.6% compared to 71.2%.

The paper also included a "sensitivity analysis," which is like checking how fussy the method is. They found that the method works best when it looks at up to 4 hops away (a resolution depth of J=3, meaning 0, 1, 2, and 4 hops) and when the "grafting strength" (how much it adjusts the data) is set to 0.5. If they adjusted the data too much or too little, the performance dropped.

The Takeaway

The paper doesn't claim to have solved the entire problem of AI adaptation, but it suggests that the old way of assuming "same distance equals same meaning" is flawed. By explicitly learning how to match different "viewing distances" across different domains, and by being gentle and probabilistic about how it applies that knowledge, CReSL manages to transfer knowledge more effectively. It's a reminder that in the world of AI, sometimes you have to look further away to see the picture clearly, and you have to be smart about which "zoom level" you use for the job.

Technical Summary: Cross-Resolution Semantic Learning for Graph Domain Adaptation

Problem Definition

The paper addresses Graph Domain Adaptation (GDA), the task of transferring predictive knowledge from a labeled source graph domain to an unlabeled target domain under distribution shifts. While existing methods focus on aligning representations or regularizing graph structures, the authors identify a critical, unaddressed limitation: Semantic Resolution Shift.

In GDA, the "propagation resolution" refers to the neighborhood range encoded by a graph representation (e.g., 0-hop for node features, 1-hop for immediate neighbors, etc.). The authors observe that the propagation resolution at which class-discriminative evidence is strongest often differs between the source and target domains. For instance, a feature learned at a 1-hop range in the source domain might be most informative at a 2-hop range in the target domain. Existing methods typically enforce implicit same-resolution transfer, assuming that knowledge learned at resolution $r$ in the source is best applied at resolution $r$ in the target. This assumption leads to suboptimal performance and negative transfer when the optimal resolutions do not align.

The core challenge is to learn a soft source-to-target resolution correspondence without target labels, while handling three specific difficulties:

Comparable Resolution-Specific Representations: Preserving distinct evidence for different resolutions within a shared latent space without erasing resolution identity.
Correspondence without Target Labels: Inferring resolution compatibility without ground truth, avoiding the "coupled estimation" problem where premature hard assignments reinforce errors.
Instance-Level Transfer under Class Uncertainty: Applying global resolution correspondence to individual target graphs whose class membership is uncertain, avoiding hard assignments that shift representations into incorrect semantic regions.

Methodology: CReSL

The authors propose Cross-Resolution Semantic Learning (CReSL), a framework designed to learn and exploit cross-resolution correspondences. The method consists of three main components:

1. Multi-Resolution Representation Bank with Source Experts

To capture evidence across different neighborhood ranges, CReSL constructs a bank of representations using a shared Graph Neural Network (GNN) encoder and learnable resolution embeddings.

Resolution Encoding: For a graph $G$ , features are propagated at multiple pre-defined orders $R = \{r_0, \dots, r_J\}$ (e.g., 0, 1, 2, 4 hops).
Shared Space: A shared encoder $\Phi_\theta$ processes these multi-resolution inputs. To maintain resolution identity, learnable embeddings $E_{res}$ are concatenated with the node features before encoding.
Source Experts: Each source resolution $j$ is associated with a lightweight linear expert $h_j$ . A global mixture weight $\alpha_j$ determines the contribution of each expert. This setup allows the model to evaluate target representations at any resolution $k$ using the expert trained on source resolution $j$ , enabling cross-resolution routing.

2. Cross-Resolution Prototype Transport (CRPT)

This module learns the soft correspondence between source and target resolutions by analyzing class-conditional structures.

Prototype Construction: Class-resolution prototypes are constructed for the source (using ground truth labels) and the target (using soft posteriors).
- Source prototype $\mu^s_{c,j}$ : Mean of source representations for class $c$ at resolution $j$ .
- Target prototype $\mu^t_{c,k}$ : Probability-weighted mean of target representations for class $c$ at resolution $k$ .
Discrepancy Measurement: The compatibility between source resolution $j$ and target resolution $k$ is measured by the class-averaged Euclidean distance between their prototypes ( $D_{j,k}$ ).
Routing Matrix: A soft routing matrix $\Gamma$ is updated via row-wise softmax normalization of the negative discrepancies. This matrix dictates how much weight source expert $h_j$ should place on target resolution $k$ .
Objective: The CRPT loss minimizes the routing-weighted prototype discrepancy, encouraging the model to route experts to compatible target resolutions.

3. Cross-Resolution Target Grafting (CRTG)

To adapt individual target graphs under class uncertainty, CRTG translates the global correspondence into instance-specific adjustments.

Prototype Displacement: For each class $c$ and resolution pair $(j, k)$ , a displacement vector $\Delta^{j \leftarrow k}_c = \mu^s_{c,j} - \mu^t_{c,k}$ is computed, representing the shift from the target prototype to the source prototype.
Posterior-Weighted Mixing: For a target graph $G^t_i$ , the displacement is a weighted sum of class-conditioned displacements based on the current predicted posterior $p^t_{i,c}$ . This avoids hard class assignment.
Grafting: The target representation at resolution $k$ is "grafted" toward the source structure by adding the weighted displacement: $\tilde{z}^{j \leftarrow k}_{G^t_i} = z^{(k)}_{G^t_i} + \beta \Delta^{j \leftarrow k}_i$ .
Consistency Objective: A KL-divergence loss enforces that the prediction from the grafted representation (via expert $h_j$ ) aligns with the aggregate target posterior, weighted by the routing matrix $\Gamma$ .

Key Contributions

Problem Formulation: The paper formally defines Semantic Resolution Shift in GDA as the cross-domain change in the resolution profile of class-discriminative evidence, identifying the limitations of fixed same-resolution transfer.
CReSL Framework: The authors propose a novel method featuring:
- A multi-resolution representation bank with resolution-specific experts.
- CRPT: A mechanism to learn soft cross-resolution correspondence using class-resolution prototypes without target labels.
- CRTG: A posterior-conditioned grafting mechanism that performs instance-level adaptation under class uncertainty.
Empirical Validation: Extensive experiments on graph classification benchmarks under diverse domain shifts (structural and feature-based) demonstrate that CReSL outperforms state-of-the-art baselines.

Experimental Results

The authors evaluated CReSL on several datasets, including Mutagenicity, PROTEINS, NCI1, ogbg-molhiv, DD, BZR, and COX2, under node/edge density shifts and feature distribution shifts.

Performance: CReSL achieved the best performance in the majority of settings, outperforming strong baselines such as GAA, TDSS, DEAL, and standard GNNs (GCN, GIN).
Ablation Studies: Removing the multi-resolution bank (CReSL w/o MR) or the learned routing (CReSL w/o CRR) significantly degraded performance, confirming the necessity of handling resolution shifts. Removing CRPT or CRTG also led to performance drops, validating their complementary roles in aligning class structures and refining instance-level representations.
Analysis:
- Correspondence: Visualizations showed that the learned routing matrix $\Gamma$ closely matched an "oracle" matrix (computed with target labels), correctly identifying that, for example, 0-hop and 1-hop source experts should route to 2-hop target resolutions.
- Grafting: t-SNE visualizations demonstrated that CRTG effectively moves dispersed target representations toward compact, source-aligned class clusters without uniform translation, preserving instance-specific nuances.

Significance and Claims

The paper claims that CReSL advances the field of Graph Domain Adaptation by explicitly modeling the mismatch in propagation resolutions between domains. Unlike previous methods that align representations at a single domain level or assume fixed resolution pairing, CReSL learns a soft, data-driven correspondence that routes source knowledge to the most compatible target resolutions.

The authors emphasize that their approach:

Avoids the negative transfer caused by forcing knowledge learned at one resolution to be applied at the same resolution in the target.
Handles class uncertainty in the target domain by using soft posteriors for both correspondence learning and instance-level grafting, rather than relying on hard pseudo-labels.
Provides a theoretical bound on target risk, showing that minimizing cross-resolution prototype discrepancy and grafting inconsistency tightens the generalization bound.

The work concludes that explicitly addressing semantic resolution shift is crucial for effective GDA and suggests future directions in extending these concepts to continuous resolution spaces.