🤖 machine learning

RHEA: Reliability-Harmonized Reconstruction and Assignment for Robust Multimodal-Attributed Graph Clustering

RHEA is a robust multimodal-attributed graph clustering framework that enhances performance under noisy or missing attributes by estimating node-specific modality reliability through neighborhood consensus to guide adaptive fusion, representation reconstruction, and topology-aware clustering.

Original authors: Yinlin Zhu, Di Wu, Ziyu Han, Zekai Chenm, Wang Luo, Miao Hu, Guocong Quan

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yinlin Zhu, Di Wu, Ziyu Han, Zekai Chenm, Wang Luo, Miao Hu, Guocong Quan

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 organize a massive, chaotic library where every book has two different cover stories: a written blurb on the back and a picture on the front. In the world of computer science, this is called a Multimodal-Attributed Graph. Think of the "graph" as a giant web of connections (like friends on social media or products bought together), and the "multimodal" part as those two different types of information (text and images) attached to every single node in that web. Scientists use these webs to group things together automatically—like finding communities of people who love the same music or sorting thousands of products into categories without a human having to read every label.

But here's the catch: in the real world, data is messy. Sometimes a book's picture is torn, blurry, or missing entirely. Sometimes the text is full of typos or nonsense. Most computer programs that try to organize these webs assume that every book's picture and text are equally perfect and trustworthy. They treat a blurry, corrupted image the same way they treat a crystal-clear one, which often leads to the whole sorting system getting confused and making mistakes. The big question researchers are trying to solve is: How can a computer figure out which pieces of information are reliable and which are junk, without anyone telling it the answers beforehand?

This is where a new method called RHEA (Reliability-Harmonized Reconstruction and Assignment) comes in. The researchers behind RHEA realized that in a connected web, your neighbors usually know what you're about. If you are a "science fiction" book, your neighbors are likely science fiction books too. So, if your picture is blurry but your neighbors' pictures are all clear and look like spaceships, the computer can guess that your picture is the problem, not the genre. RHEA uses this "neighborly gossip" to figure out which data is trustworthy and which is broken.

Instead of blindly trusting every piece of data, RHEA acts like a smart librarian who checks the crowd before making a decision. It looks at a node (a book) and asks, "Does your text match your neighbors? Does your picture match them?" If a node's data doesn't fit with the group, RHEA flags it as unreliable. Then, it does something clever: it doesn't just throw that bad data away. Instead, it "reconstructs" it by borrowing the clear, reliable information from the trustworthy neighbors. It essentially says, "Since your picture is ruined, let's use the average of your neighbors' pictures to guess what yours should look like."

Once the data is cleaned up and the unreliable bits are fixed, RHEA uses a special mathematical tool called "optimal transport" to sort everything into groups. Think of this like moving furniture into rooms; RHEA makes sure that the heavy, reliable pieces of information (the clear pictures and text) carry more weight in deciding which room a book belongs in, while the lighter, reconstructed pieces carry less weight. This ensures that the final groups are accurate even if some of the original data was terrible.

The researchers tested RHEA on four different real-world datasets, including social networks and e-commerce catalogs, under five different conditions ranging from perfect data to heavily corrupted data. They found that RHEA consistently outperformed the best existing methods. The more messy the data got, the bigger the advantage RHEA had. In fact, when they artificially corrupted the data to test it, RHEA was able to detect the corruption with over 95% accuracy, proving that its "neighborly gossip" method is a very effective way to spot and fix bad information. By learning to trust the crowd and repair the broken parts, RHEA makes it possible to organize complex, messy data much more reliably than ever before.

Technical Summary: RHEA – Reliability-Harmonized Reconstruction and Assignment for Robust Multimodal-Attributed Graph Clustering

1. Problem Statement

Multimodal-attributed graphs (MAGs) represent nodes with heterogeneous attributes (e.g., text and images) over a relational structure, serving as a foundation for tasks like community discovery and product segmentation. While existing multimodal graph clustering methods perform well when attributes are clean and complete, they degrade significantly under real-world conditions where attributes are noisy, corrupted, or missing.

The core limitation of current approaches is their implicit assumption that modality reliability is uniform across all nodes. In practice, reliability is node-specific: an image may be corrupted for one node while its text is incomplete, or vice versa. Existing methods fail to distinguish between reliable and unreliable modalities at the individual node level, leading to fused representations contaminated by noise or unable to compensate for missing data. Furthermore, in unsupervised settings, there are no labels to infer this reliability, posing a fundamental challenge: how to estimate node-specific modality reliability without supervision.

2. Methodology: RHEA

The authors propose RHEA (Reliability-Harmonized rEconstruction and Assignment), a framework that leverages the principle of attribute homophily—the observation that neighboring nodes tend to share similar semantic attributes—to estimate reliability and guide clustering.

2.1 Neighbor-Consensus Reliability Estimation

RHEA estimates a node-specific modality reliability field $\rho_{i,m} \in [0, 1]$ in a fully unsupervised manner.

Mechanism: For each node $i$ and modality $m$ , the method computes an intra-modal neighbor agreement score $a_{i,m}$ based on the cosine similarity between the node's embedding and its neighbors' embeddings in that modality.
Reliability Field: These agreement scores are converted into a probability distribution over modalities for each node using a softmax function with a temperature parameter $\tau$ . A modality consistent with its neighborhood is assigned high reliability; one that deviates is assigned low reliability.

2.2 Reliability-Aware Representation Learning

Once reliability is estimated, RHEA reconstructs and fuses representations adaptively:

Reconstruction: If a modality is missing or its reliability score falls below a threshold $\gamma$ , the representation is reconstructed by aggregating embeddings from neighboring nodes that have observed, reliable data. This prevents the propagation of missing or corrupted signals.
Confidence Estimation: A reconstruction confidence score $c_i$ is calculated based on the number of available neighbors and the variance of their embeddings. This serves as an absolute reliability measure.
Adaptive Fusion: The final fused embedding $z_i$ is a weighted sum of modality-specific embeddings, where weights are determined by the estimated reliability $\rho_{i,m}$ . If a modality is reconstructed, the reconstructed embedding is used; otherwise, the original observed embedding is retained.

2.3 Reliability-Aware Clustering

The clustering stage utilizes the reliability information to optimize cluster assignments via Optimal Transport (OT) and Assignment Distillation:

Reliability-Aware OT: The transport cost between node embeddings and cluster prototypes is computed. Crucially, the row marginals of the transport plan are weighted by the reconstruction confidence $c_i$ . This ensures that nodes with uncertain reconstructions contribute less mass to the transport plan, reducing their influence on the final assignment.
Neighbor-Consensus Assignment Distillation (NCRC): To further regularize assignments, a soft cluster assignment is propagated over the graph for $L$ hops. This neighbor-consensus target is sharpened and used as a distillation objective to encourage locally consistent cluster assignments, particularly for nodes where direct evidence is weak.

2.4 Optimization

The model is trained end-to-end by minimizing a joint objective function comprising:

Clustering Loss ( $L_{clu}$ ): The optimal transport cost.
Distillation Loss ( $L_{ncrc}$ ): The neighbor-consensus assignment loss.
Auxiliary Contrastive Losses: A graph-walk contrastive loss ( $L_{nbr}$ ) to preserve topology-aware invariance and a cross-modal alignment loss ( $L_{mod}$ ) to align modality-specific representations with the fused embedding.

3. Key Contributions

The paper identifies three primary contributions:

Valuable Insights: The authors identify node-specific modality reliability as a missing component for robust MAG clustering. They demonstrate that graph neighborhoods provide supervision-free evidence for estimating this reliability under attribute homophily.
Novel Method: RHEA is proposed as a unified framework where a shared neighborhood-consensus reliability signal simultaneously guides:
- Graph-based modality reconstruction.
- Adaptive multimodal fusion.
- Topology-aware optimal transport clustering.
State-of-the-Art Performance: Extensive experiments show consistent improvements over existing methods, with gains becoming more pronounced as attribute quality deteriorates.

4. Experimental Results

RHEA was evaluated on four MAG benchmarks (RedditS, Toys, Grocery, Amazon) under five attribute conditions: standard (clean), image corruption, image missingness, text corruption, and text missingness.

Standard Setting: RHEA achieved the best performance across all four datasets and all metrics (NMI, ACC, ARI, F1), outperforming strong baselines like DGF and DMGC.
Robustness to Perturbation: Under corrupted or missing modalities, RHEA consistently outperformed all baselines. The performance gap widened as perturbation severity increased. For instance, on the RedditS dataset with 80% image missingness, RHEA maintained an NMI of 0.729, while the next best method (DGF) dropped to 0.647.
Reliability Detection: The learned reliability field was shown to accurately detect synthetically injected corruption, achieving an AUROC consistently exceeding 0.95 across all datasets.
Ablation Studies: Removing the reconstruction module caused the largest performance drop under missingness, while removing the reliability gating caused the largest drop under corruption. This confirms that reconstruction handles missing data, while gating prevents noise injection in corrupted data.

5. Significance and Claims

The paper claims that RHEA demonstrates the importance of modeling node-specific, graph-induced reliability for robust multimodal representation learning. By treating reliability as a latent variable that varies across nodes and modalities, rather than a global property, the framework effectively repairs unreliable observations using local structural context.

The authors position RHEA as a solution to the "clean data assumption" prevalent in current MAG clustering methods. They argue that local structural context provides an effective signal for identifying unreliable observations, enabling robust clustering even when data quality is heterogeneous. The paper concludes by noting that future work will investigate extending the framework to low-homophily and heterophilous graphs, acknowledging that the current method relies on the assumption that neighbors share similar attributes.