acceptodds
Withdrawn from ICLR 2027

FedN-TAGC: Federated Node-Level Text-Attributed Graph Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Node-level federated graph clustering studies how several holders of unlabeled subgraphs can train clustering models without sharing local graph data. Existing methods usually rely on graph topology or vector attributes, but many real graphs also attach texts to nodes. In text-attributed graphs, a node's text and its local neighborhood may point to different clusters, which can corrupt local cluster evidence and cross-client semantic alignment. We propose FedN-TAGC, a Federated Node-Level Text-Attributed Graph Clustering method for this setting. On each client, FedN-TAGC learns text-view, structure-view, and fused node representations, and compares text-view and structure-view assignments under the same fused cluster centers to separate stable samples from structure-text conflict cases. Stable samples are summarized by an LLM into local cluster descriptions, while conflict cases are assigned after the server forms global topics. The server groups uploaded summaries into global topics, aligns client clusters, and sends semantic consensus back to guide later local clustering. Experiments on benchmark text-attributed graphs show that this conflict-handling design improves federated node clustering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Federated Graph Clustering, Node-Level Clustering, Unsupervised Graph Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_cdGWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the problem of federated graph clustering, focusing on the disagreement between local conflict screening and global semantic consensus. The authors first select reliable nodes to generate high-quality local summaries, which are then passed to the global consensus stage. Extensive experiments on multiple datasets with varying numbers of clients have demonstrated the effectiveness of the proposed framework. Despite these merits, several concerns still need to be clarified.

Critical strengths

1.The paper is well organized and clearly written.

2.The problem of federated graph clustering is interesting and meaningful.

3.The experiments are extensive and helpful to validate the effectiveness of the framework.

Critical weaknesses

1.In Eq. (11), the node features are fixed embeddings, and the normalized embedding in local cluster k depends on the cluster assignments. Additionally, the term \phi also appears to be fixed. Although the set of stable nodes varies according to the K-means clustering results, the paper employs discrete cluster assignments and node selection without providing a differentiable alternative mechanism. Consequently, based on the current formulation, it is unclear how this global semantic loss can directly optimize the local GNN encoders.

2.Several baselines do not incorporate the reasoning capabilities of LLMs. Therefore, it is necessary to distinguish the contribution of the proposed methodology from that of the LLMs.

3.One of the main motivations for federated graph clustering is privacy preservation, while the use of LLMs seems to require representative textual data. How does the proposed method properly address privacy concerns and computational efficiency? Additionally, an efficiency analysis of this method is required.

4.Evaluating the generalizability of the proposed method to larger-scale datasets and more clients would be valuable.

Questions for authors

Please answer the above weaknesses.

Reviewer_fWJ5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated node-level clustering on text-attributed graphs, where local graph structure and node text may provide conflicting clustering signals. The authors propose FedN-TAGC, which detects text-structure misalignment at each client, summarizes stable local cluster evidence with an LLM, and constructs a server-side global semantic topic space for cross-client cluster alignment and ambiguous-node calibration. Experiments on five text-attributed graph benchmarks show consistent improvements over federated graph clustering and adapted federated graph learning baselines.

Critical strengths
  1. The paper studies an interesting and relatively underexplored problem. Federated graph clustering has been studied previously. The paper identifies a meaningful failure mode: text and structure may support different latent topics. Directly summarizing such mixed evidence can propagate local errors into global cluster alignment. This provides a clear motivation for the proposed conflict-aware framework.
  2. The overall design is coherent and well matched to the problem setting. Comparing the text-only and structure-propagated views under the same cluster centers provides a common reference for detecting cross-view disagreement. Delaying ambiguous-node assignment until a global semantic topic space has been constructed is also reasonable.
  3. The empirical evaluation is substantial, and the performance is plausible. The experiments cover five datasets and three client scales, and FedN-TAGC consistently achieves clear improvements over the considered baselines across ACC, NMI, ARI, and macro-F1.
Critical weaknesses
  1. The assumption that text-structure disagreement indicates unreliable evidence is not sufficiently justified. In heterophilic graphs, graph structure may legitimately encode information different from the node text, and such disagreement may itself be informative rather than noise.
  2. The method relies heavily on a powerful external LLM, and the contribution of the proposed algorithmic design versus the LLM itself is not fully isolated. GPT-5 is used for local cluster summarization, global topic construction/alignment, and ambiguous-node semantic resolution. These operations are central to the proposed framework. It is therefore unclear how much of the performance gain comes from the proposed conflict-handling mechanism or from access to a very strong semantic model.
  3. The experimental protocol relies on knowing the ground-truth number of classes K. In a truly unsupervised federated clustering scenario, the correct number of global topics is generally unknown. Consequently, the current setting may substantially simplify both local clustering and global topic alignment.
Questions for authors
  1. Can the authors evaluate cases where text and graph structure are intentionally complementary rather than one view being unreliable?
  2. How strongly does the proposed conflict score correlate with actual clustering errors? Can the authors report the clustering accuracy or confidence of the detected conflict nodes versus the stable nodes using labels only for post-hoc analysis?
  3. How sensitive is FedN-TAGC to the choice of LLM? For example, what happens when GPT-5 is replaced by a smaller open-source model or a weaker language model?
  4. How does the method perform when the true number of clusters is unknown? Can the server infer or dynamically adapt the number of global topics instead of fixing it to the number of ground-truth classes?
Reviewer_hphcOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a model called FedN-TAGC to address structure–text conflicts and cross-client semantic misalignment in unsupervised federated graph clustering, which uses LLM-generated summaries of reliable nodes to establish global semantic consensus and resolves uncertain assignments through this consensus.

Critical strengths

The research problem is interesting and important in unsupervised federated graph clustering.

Critical weaknesses
  1. Lack of novelty: both the framework and the used components in the framework are not original, making this paper looks more like an engineering paper instead of a research paper.

  2. Risk of data leakage: the local cluster summaries generated by LLM in LSRC may contain data information in local clients, which should not be sent to the global server following the standard application of federated learning.

Questions for authors
  1. What is the original contribution of your paper?

  2. How to prevent the potential data leakage by using the local cluster summaries in LSRC?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.