# Node Role-Guided LLMs for Dynamic Graph Clustering

Source: [Node Role-Guided LLMs for Dynamic Graph Clustering](<https://www.sota2.com/research/paper/69b8dfed0636e5bf4afbe871>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-18T05:00:04.467Z

Authors: Dongyuan Li, Ying Zhang, Yaozu Wu, Renhe Jiang

Published: 2026-03-14T07:14:53.000Z

## Abstract

Dynamic graph clustering aims to detect and track time-varying clusters in dynamic graphs, revealing how complex real-world systems evolve over time. However, existing methods are predominantly black-box models. They lack interpretability in their clustering decisions and fail to provide semantic explanations of why clusters form or how they evolve, severely limiting their use in safety-critical domains such as healthcare or transportation. To address these limitations, we propose an end-to-end interpretable framework that maps continuous graph embeddings into discrete semantic concepts through learnable prototypes. Specifically, we first decompose node representations into orthogonal role and clustering subspaces, so that nodes with similar roles (e.g., hubs, bridges) but different cluster affiliations can be properly distinguished. We then introduce five node role prototypes (Leader, Contributor, Wanderer, Connector, Newcomer) in the role subspace as semantic anchors, transforming continuous embeddings into discrete concepts to facilitate LLM understanding of node roles within communities. Finally, we design a hierarchical LLM reasoning mechanism to generate both clustering results and natural language explanations, while providing consistency feedback as weak supervision to refine node representations. Experimental results on four synthetic and six real-world benchmarks demonstrate the effectiveness, interpretability, and robustness of DyG-RoLLM. Code is available at https://github.com/Clearloveyuan/DyG-RoLLM.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Graph Clustering on EC-30K](<https://www.sota2.com/research/sota/graph-clustering-on-ec-30k.md>): Graph Clustering; EC-30K; NMI: 99.4.

- [Graph Clustering on Wikipedia](<https://www.sota2.com/research/sota/graph-clustering-on-wikipedia.md>): Graph Clustering; Wikipedia; NMI: 0.516.

- [Graph Clustering on Dublin](<https://www.sota2.com/research/sota/graph-clustering-on-dublin.md>): Graph Clustering; Dublin; NMI: 57.6.

- [Graph Clustering on arXiv](<https://www.sota2.com/research/sota/graph-clustering-on-arxiv.md>): Graph Clustering; arXiv; NMI: 53.3.

- [Graph Clustering on BD-100K](<https://www.sota2.com/research/sota/graph-clustering-on-bd-100k.md>): Graph Clustering; BD-100K; NMI: 94.3.

- [Graph Clustering on DR 100K](<https://www.sota2.com/research/sota/graph-clustering-on-dr-100k.md>): Graph Clustering; DR 100K; NMI: 96.2.

- [Graph Clustering on MS 100K](<https://www.sota2.com/research/sota/graph-clustering-on-ms-100k.md>): Graph Clustering; MS 100K; NMI: 95.4.

- [Graph Clustering on SO](<https://www.sota2.com/research/sota/graph-clustering-on-so.md>): Graph Clustering; SO; NMI: 47.6.

- [Graph Clustering on Flickr](<https://www.sota2.com/research/sota/graph-clustering-on-flickr.md>): Graph Clustering; Flickr; NMI: 54.2.

- [Graph Clustering on Youtube](<https://www.sota2.com/research/sota/graph-clustering-on-youtube.md>): Graph Clustering; YouTube; NMI: 52.7.

- [Graph Clustering on BD-30K](<https://www.sota2.com/research/sota/graph-clustering-on-bd-30k.md>): Graph Clustering; BD-30K; NMI: 99.9.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.13799>)

[PDF](<https://arxiv.org/pdf/2603.13799.pdf>)
