acceptodds
Under review as a conference paper at ICLR 2027

CSRI-FGL: Cross-Client Semantic Relation Inference for Federated Graph Learning

Abstract

Federated graph learning (FGL) enables collaborative learning over distributed graphs while keeping raw graph data local. However, fragmented and non-IID local subgraphs make it difficult for existing parameter-aggregation-based methods to capture transferable semantic connection patterns for recovering semantically meaningful missing links. To address this issue, we propose CSRI-FGL, a framework that learns semantic connection patterns for missing-link recovery via LLM-guided relation inference. It distills node-level textual semantics and pairwise textual consistency signals into federatively aggregated edge generators, enabling each client to infer informative links without sharing raw graph data. Specifically, CSRI-FGL first extracts semantic representations and textual consistency signals from local node texts using a client-side LLM module to supervise the construction of local-pattern-enhanced graphs. The server then aggregates client-side edge-generator parameters into a global generator that captures transferable connection patterns across clients, and each client applies this global generator only within its own node set to build a global-pattern-enhanced local graph. Finally, a dual-branch GNN decouples personalized node representation learning from shared semantic connection-pattern learning by separately modeling local-pattern-enhanced and global-pattern-enhanced graphs. Experiments on five node-classification benchmarks under non-IID client partitions demonstrate that CSRI-FGL consistently outperforms representative FGL baselines in overall node-classification performance while maintaining competitive personalized performance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Federated Graph Learning, Graph Neural Networks, Large Language Models, Graph Structure Learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_hNAdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies federated node classification on text-attributed graphs under non-IID client partitions. The authors propose CSRI-FGL, which uses a frozen client-side LLM to extract node-level semantic representations and pairwise textual consistency scores to supervise local semantic edge generators, then obtains a global semantic edge generator through federated aggregation, applies the local and global generators to add edges within each client's own node set, and finally uses a dual-branch GraphSAGE architecture to process the two enhanced graphs and fuse their representations with textual semantic representations for node classification. The method does not generate cross-client edges. Experiments show that, in the reported mean macro-F1 results, CSRI-FGL outperforms the compared baselines on Collaborative Test across all five datasets while maintaining competitive Local Test performance.

Critical strengths
  1. The method shifts the object of collaboration from concrete node relations to reusable connection functions. It distills semantic supervision from local node pairs into aggregatable semantic edge generators, enabling other clients to apply the learned rules to their own node pairs without exchanging node-level representations, aligning nodes across clients, or constructing a server-side global graph.
  2. The method provides a targeted mechanism to mitigate potential negative transfer caused by shared structural priors. The two branches perform message passing separately on different enhanced graphs, avoiding the merging of the two sets of added edges into a single neighborhood beforehand; the branch-level consistency term uses stop-gradient local representations as a reference to guide the global branch toward alignment. In Table 4, the variant using global-pattern enhancement without dual-branch decoupling performs worse on Collaborative Test than the local-pattern enhancement variant, whereas the full model performs better, which is consistent with the motivation of preserving a local structural view while leveraging shared relation knowledge.
Critical weaknesses
  1. The ablations do not sufficiently isolate the benefit of federated edge-generator aggregation. In Table 4, global w/o Dual-branch underperforms both + Local-pattern and Baseline on Collaborative Test. The full model also changes the branch architecture, leaving the contribution of shared connection patterns unclear. A matched dual-branch control without edge-generator aggregation would clarify this contribution.
  2. Baseline adaptations under the common information constraints remain unclear. FedEgo uploads mashed ego-graphs, while DFGL2GS uses client-generated graphs for server-side learning. The main paper does not clarify whether these transmissions and core mechanisms are retained.
Questions for authors
  1. With semantic inputs, dual-branch architecture, edge budgets, and federated GNN training fixed, how does disabling edge-generator aggregation affect performance, and how is its regularization handled?
  2. In Collaborative Test, how are reserved nodes connected and predictions produced or aggregated? Are the evaluation rules and accessible information consistent across baselines?
2 comments
AuthorsOpenReview ↗

Author Response to Reviewer hNAd

We thank the reviewer for recognizing the motivation for transferable semantic relation functions and the dual-branch design. We address both concerns below.

W1Q1: Independent contribution of federated edge-generator aggregation.

The original Table 4 ablations change both graph-enhancement components and branch configurations, so they do not independently isolate generator aggregation. We therefore introduced Full w/o Generator Aggregation within the complete dual-branch framework. In this variant, server-side aggregation and redistribution of edge-generator parameters are disabled, while each client independently retains and updates its local generator. The Local-pattern branch remains unchanged, and the Global-pattern branch uses the corresponding client-local generator instead of the aggregated generator. Both GraphSAGE branches, candidate-pair scoring, structural-semantic fusion, threshold/top-κ selection, edge budgets, and downstream training settings are retained. The asymmetric consistency objective also retains its original coefficient and stop-gradient direction. Under the original Table 4 setting (K = 6; Citeseer, PubMed, and Wiki), Full w/o Generator Aggregation achieves a mean Collaborative Test Macro-F1 of 0.7936, compared with 0.8193 for Full CSRI-FGL, corresponding to an improvement of 2.57 percentage points. These results demonstrate the empirical benefit of federated semantic relation-function sharing over independently trained client-local generators under the evaluated setting. Our contribution lies in the federated sharing and application of transferable semantic relation functions, rather than proposing a new FedAvg algorithm.

W2Q2: Collaborative Test protocol and baseline fidelity.

Collaborative Test. Before non-IID client partitioning, approximately 30% of nodes are reserved by class-stratified sampling. These nodes and all their original incident edges are removed from the training graphs. They do not participate in local training, validation, Local Test, training-time message passing, edge-generator optimization, or model selection. The remaining nodes are partitioned across clients and split 80%/10%/10% locally (Section 4.1). At inference, each reserved node is assigned to exactly one participating client without using its ground-truth label for client assignment, and temporarily inserted into that client's local evaluation graph. Its original features and cached Mistral semantic representation are available, but its original incident edges remain fully masked. CSRI-FGL may generate semantic links between reserved and existing local nodes or between reserved nodes, strictly within the assigned client. No cross-client edges, neighbor queries, node-level representation sharing, or raw-graph exchange are involved in CSRI-FGL inference. Predictions are evaluated only for reserved nodes. We pool these predictions across all clients and compute one Collaborative Macro-F1 over the combined reserved set rather than averaging client-level scores. Ground-truth labels are used for the initial class-stratified sampling and final metric computation, not for client assignment, reserved-node training, or model selection. Thus, Collaborative Test evaluates previously unseen nodes within participating clients, not zero-shot generalization to entirely unseen clients. Baseline fidelity. FedEgo retains its mashed ego-graph upload mechanism, whereas DFGL2GS retains client-generated graph uploads and server-side graph learning. Neither mechanism is replaced with CSRI-FGL's generator-parameter sharing. We keep datasets, client partitions, evaluation splits, and metrics comparable while preserving each baseline's own collaboration and learning procedures.

Thank you for the detailed response. The proposed experiments and analyses address several of my concerns. I will retain my current rating.

Reviewer_uwFrOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

CSRI-FGL learns client-local semantic edge generators from observed graph connectivity and frozen LLM-derived text signals, aggregates generator parameters across clients, and constructs local-pattern and global-pattern enhanced views for dual-branch GraphSAGE classification. The paper studies personalized and collaborative prediction on five text-attributed graph benchmarks. The main methodological question is how textual relation evidence can be learned locally and reused across clients without explicitly building a cross-client graph.

Critical strengths
  1. The edge generator uses textual evidence without treating semantic similarity as observed adjacency. Equation (14) combines connectivity labels with soft LLM pairwise targets to train a reusable relation scorer. This matters for fragmented client graphs: existing edges may miss useful relationships, while high text similarity alone is an unreliable rule for adding GNN neighbors. The teacher signal also has empirical support. Supplementary Table S2 reports Mistral pair-score AUROC of 0.8643 versus 0.8361 for raw TF-IDF cosine on sampled pairs. The comparison supports learning from richer pairwise targets rather than relying on an untrained similarity threshold.
  2. Graph enhancement uses a selective structural–semantic rule instead of adding every plausible text-based connection. Equations (5)–(8) consider candidate non-edges within each client, combine learned semantic scores with structural proximity, and apply top-κ selection with a threshold. This design matters because changing a GNN's neighbors changes the information propagated during classification. Structural evidence moderates semantic suggestions, while the selection budget limits the number of inferred connections. The procedure therefore specifies how relation scores become usable graph edges, rather than leaving augmentation to an unconstrained similarity graph.
  3. Parameter aggregation makes cross-client relation sharing possible without constructing a global adjacency matrix. Equation (9) averages locally trained generators; the resulting model scores candidate pairs only within each receiving client's graph. Client node identities and edge sets need not be aligned for this exchange, because the shared object is a relation-scoring function rather than a list of connections. This distinction gives the federation step a concrete role: knowledge learned from other clients can inform local graph construction while node texts and generated edge sets remain at their respective clients.
  4. Separate graph views preserve the distinction between client-specific and globally guided connections during classification. The two GraphSAGE branches process these neighborhoods independently, avoiding the loss of their provenance through early edge merging. Equation (13) introduces a stop-gradient consistency term that coordinates branch representations while keeping their graph inputs distinct. This is an appropriate design choice for heterogeneous clients: shared relation patterns can influence prediction without replacing locally observed structural context. The architecture thus addresses how to combine the two sources of graph information, rather than merely increasing the number of inferred edges.
Critical weaknesses
  1. The supervision scheme and the interpretation of recovered edges need a tighter connection. Under Equation (14), a sampled non-edge can carry a negative BCE label even when its LLM consistency score is high; the effect of this disagreement on learned connection probabilities is not examined. Moreover, Supplementary Table S2 measures standalone pair-score discrimination, whereas actual new edges are chosen through the combined structural–semantic top-κ procedure. Candidate-level AUROC does not establish precision among selected connections. A fixed-budget audit of disagreement cases and final selected edges against held-out or independently verified relations would strengthen the relation-recovery claim.
  2. Parameter aggregation does not by itself verify transfer to previously unseen client distributions. Equation (9) shares generators among participating clients, while the main evaluations measure downstream performance on the evaluated client populations. Their gains could arise from patterns useful within those populations without showing adaptation to different text or class distributions. A held-out-client comparison of global and locally trained generators would clarify the extent of transferable relation learning.
  3. The direction of the asymmetric consistency objective in Eq. (13) deserves a more direct test. The globally guided branch is aligned toward a stopped-gradient local representation, making client-specific features the reference for consistency. Under heterogeneous client graphs, this choice could also carry local structural or class biases into the global branch. The full-model ablation does not isolate this direction from the overall benefit of using two enhanced views. Comparing reversed or symmetric alignment while holding graph views and model capacity fixed would clarify what the chosen direction contributes.
Questions for authors

Please refer to the Weaknesses section.

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer uwFr

We thank the reviewer for the positive assessment of semantic supervision, federated relation sharing, and the dual-branch design. We respond to the three concerns directly.

W1Q1: Conflicting supervision and reliability of selected semantic edges.

We agree that the two supervision signals in Eq. (14) can disagree: a sampled non-edge has a negative observed-adjacency label but may receive a high LLM semantic-consistency score. This is not inherently contradictory because observed graph connectivity and semantic relatedness encode different information. A high semantic score alone does not establish the existence of an original or missing ground-truth edge. The generator is trained with binary cross-entropy on locally observed edges and sampled non-edges together with LLM-derived soft pairwise targets. At graph-enhancement time, its score is combined with local structural proximity (Eq. (6)); candidate non-edges are then subject to a threshold and per-node top-κ selection (Eq. (7)). Thus, high semantic affinity does not automatically yield an augmented edge. Supplementary Table S2 reports candidate-pair discrimination on PubMed hard negatives at K = 6. Mistral scores achieve AUROC 0.8643 ± 0.0076, AUPRC 0.8497 ± 0.0106, and Precision@positive-count 0.7935 ± 0.0085, compared with 0.8361 ± 0.0129, 0.8386 ± 0.0158, and 0.7680 ± 0.0100 for TF-IDF cosine. These metrics evaluate standalone candidate-pair scoring, not the precision or recall of the edges finally selected after fusion and top-κ filtering. We do not claim that every added semantic edge recovers a true original edge, nor that the reported pair-scoring results directly validate the final edge set. We will state this evaluation boundary explicitly.

W2Q2: Transfer to previously unseen clients.

CSRI-FGL shares the parameters of locally trained semantic edge generators through federated aggregation, rather than exchanging node identities, explicit cross-client edges, or node-level representations. The aggregated relation-scoring function can be applied to candidate pairs in a client-local node set without cross-client node alignment (Eq. (9)). This is an architectural property, not by itself evidence of zero-shot generalization. Our current Collaborative Test withholds approximately 30% of nodes before client partitioning, removes their original incident edges, and evaluates them within participating clients without using ground-truth labels for client assignment. The original incident edges remain masked during inference. It is therefore an unseen-node evaluation, not a leave-one-client-out experiment in which an entire client is excluded from training and aggregation. The observed results support knowledge sharing among participating clients; they do not establish performance on an unseen client distribution. We will narrow the transferability claim accordingly and avoid presenting Collaborative Test as proof of zero-shot unseen-client adaptation.

W3Q3: Direction of the asymmetric consistency objective.

Equation (13) uses ||h_global − sg(h_local)||², where sg denotes stop-gradient. Under this consistency term, the global-pattern representation is aligned toward the client-specific local-pattern reference; gradients from the term do not pass through the local target. The local branch nevertheless remains trainable through its other objectives. This asymmetry is motivated by non-IID graph partitions: global relation patterns learned through federated generator aggregation may not fully reflect an individual client’s local structure. Keeping the local representation as the reference allows the global branch to absorb shared relational knowledge while discouraging the consistency loss from directly pulling the local reference toward potentially mismatched global patterns. We emphasize that this is a local-structure-preserving design rationale, not an experimentally established optimal gradient direction. The current dual-branch ablations do not independently compare the proposed direction with reversed or symmetric consistency objectives. We will state the rationale and its empirical scope explicitly, without claiming unmeasured superiority over those alternatives.

Reviewer_5EvZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a federated graph learning framework that uses LLM-derived semantic information to improve local graph structures. The method trains semantic edge generators on individual clients and aggregates their parameters to share knowledge across clients. A dual-branch GNN then combines local and global graph representations for node classification. Experiments on five benchmark datasets demonstrate improved collaborative performance while maintaining competitive local performance.

Critical strengths

The paper addresses a challenge in federated graph learning by exploring how semantic knowledge can be shared across clients without exchanging raw graph data.

The proposed framework is reasonably well organized, which employs LLM-based semantic information, federated edge generation, and dual-branch GNNs to improve local graph representations.

The experiments covering five benchmark datasets with different numbers of clients provide a comprehensive evaluation under non-IID settings.

Critical weaknesses

The paper aims to recover missing connections caused by graph partitioning, but the proposed method only adds edges between nodes within the same client. It cannot directly recover connections lost across clients.

The proposed transfer of semantic knowledge across clients mainly relies on standard FedAvg applied to edge generators. The novelty of this mechanism is relatively limited, and it remains unclear why averaging generators trained on different non-IID graphs would lead to more reliable edge predictions.

The performance improvements may largely come from the additional LLM features rather than the proposed relation inference mechanism. Since these features are directly used for classification but are unavailable to most baselines, the comparison may not fully isolate the contribution of the proposed method.

The experimental comparisons lack simple semantic graph construction baselines. For example, edges could be directly generated using cosine similarity or KNN based on LLM embeddings.

The source and preprocessing of node texts are not sufficiently explained. Some of the benchmark datasets are commonly provided with numerical or bag-of-words features rather than raw text. Since the method heavily relies on LLM-derived semantic information, the authors should clarify how these textual inputs are obtained and processed.

The paper reserves 30% of nodes before client partitioning and excludes them from training-time message passing, but it does not clearly explain how these nodes are handled during evaluation. The authors should clarify how their features and graph connections are accessed under the federated setting.

Questions for authors

The paper aims to recover missing connections caused by graph partitioning, but the proposed method only adds edges between nodes within the same client. It cannot directly recover connections lost across clients.

The proposed transfer of semantic knowledge across clients mainly relies on standard FedAvg applied to edge generators. The novelty of this mechanism is relatively limited, and it remains unclear why averaging generators trained on different non-IID graphs would lead to more reliable edge predictions.

The performance improvements may largely come from the additional LLM features rather than the proposed relation inference mechanism. Since these features are directly used for classification but are unavailable to most baselines, the comparison may not fully isolate the contribution of the proposed method.

The experimental comparisons lack simple semantic graph construction baselines. For example, edges could be directly generated using cosine similarity or KNN based on LLM embeddings.

The source and preprocessing of node texts are not sufficiently explained. Some of the benchmark datasets are commonly provided with numerical or bag-of-words features rather than raw text. Since the method heavily relies on LLM-derived semantic information, the authors should clarify how these textual inputs are obtained and processed.

The paper reserves 30% of nodes before client partitioning and excludes them from training-time message passing, but it does not clearly explain how these nodes are handled during evaluation. The authors should clarify how their features and graph connections are accessed under the federated setting.

1 comment
AuthorsOpenReview ↗

Author Response to Reviewer 5EvZ

We thank the reviewer for the constructive comments. We address all six concerns below.

W1. Cross-client connections. CSRI-FGL does not recover original cross-client edges or reconstruct the global topology. Each client trains a semantic relation scorer using local node pairs; FedAvg shares the scorer's parameters, not raw graphs, node representations, or cross-client links. The aggregated scorer constructs complementary within-client semantic edges to mitigate lost relational context. No direct cross-client message passing is restored, and generated edges are not claimed to match missing ground-truth edges.

W2. Benefit of generator aggregation. Our novelty is federated sharing of semantic relation functions, not FedAvg itself. We added Full w/o Generator Aggregation: clients retain independently trained generators; the Global-pattern branch uses the corresponding local generator instead of an aggregated one. The Local-pattern branch, two GraphSAGE encoders, candidate scoring, structural-semantic fusion, threshold/top-k selection, edge budget, and downstream training are retained. The asymmetric consistency term keeps its weight and stop-gradient direction; it is distinct from parameter-based FedReg. Under Table 4's setting (K=6; Citeseer, PubMed, Wiki), this ablation scores 0.7936 versus 0.8193 for Full (a 2.57-percentage-point gain). We will clarify the generator-related regularization treatment in this ablation so the comparison can be reproduced precisely.

W3. LLM features versus graph learning. LLM representations enter the classifier directly, so we do not attribute all gains to graph augmentation. Table 4 reports mean Collaborative Macro-F1 of 0.7390 (Baseline), 0.7590 (+LLM only), 0.7680 (+Local-pattern), and 0.8193 (Full). Supplementary Table S7 reports Full versus +LLM-only as 0.742/0.682 (Citeseer), 0.893/0.850 (PubMed), and 0.823/0.746 (Wiki). Thus LLM features contribute, while the complete framework provides additional gains from graph enhancement and dual-branch learning jointly.

W4. Simple graph-construction baselines. Under the same three-dataset K=6 setting, LLM+Cosine = 0.8129, LLM+KNN = 0.8059, and Full = 0.8193 in mean Collaborative Macro-F1. The comparisons use matched added-edge budgets and downstream training settings. Full exceeds Cosine by 0.64 and KNN by 1.34 percentage points.

W5. Text provenance. For Citeseer, paper metadata (available titles, abstracts, keywords, and content) is normalized into a Title-Description record. For PubMed, node texts are constructed by directly concatenating each paper's title (TI) and abstract (AB), without LLM-based rewriting. For Wiki, an encyclopedic prompt is designed to form Title-Description text from supplied article titles, summaries, and content. For Amazon-Computers/Photo, product titles, brands, descriptions, specifications, features, and reviews are the intended inputs to dataset-specific normalization prompts. These prompts are designed not to incorporate class labels, graph links, node IDs, or classification outputs.

W6. Reserved-node evaluation. Before partitioning clients, we class-stratify and reserve about 30% of nodes. These nodes and all original incident edges are excluded from training graphs, local train/validation/test splits, training-time message passing, generator training, and model selection. The remaining nodes are partitioned across clients and split 80/10/10 locally. For Collaborative Test, each reserved node is assigned once to a participating client without using its ground-truth label for client assignment. It is temporarily inserted into that client's evaluation graph with its original features and cached Mistral embedding, but none of its original edges are restored. CSRI-FGL may generate only within-client reserved-to-local and reserved-to-reserved semantic edges; no cross-client edges are formed. We evaluate reserved-node predictions only, pool them across clients, and compute one Macro-F1 over the combined set (not an average of client scores). This tests inductive classification of unseen nodes under masked original adjacency, not recovery of actual cross-client edges or transfer to unseen clients.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.