From Euclidean to Graph-Structured Data: A Survey of Collaborative Learning
Rémi Bourgerie, Šarūnas Girdzijauskas, Viktoria Fodor
cs.LG · cs.MA · cs.SI
Quick verdict
TMLR survey unifying federated/decentralized learning from Euclidean to graph data: taxonomy, unified algorithms, open challenges; a valuable field map.
Summary
This TMLR-published survey consolidates collaborative learning—federated and decentralized learning—across Euclidean data (images, text) and graph-structured data. Section 2 reviews the Euclidean foundations: the FedAvg framework, heterogeneity sources (partition types, statistical shift, system constraints, privacy vulnerabilities), and solutions along three axes—effectiveness (regularization, augmentation, client selection, vertical FL with collaborative inference), efficiency (aggregation topologies, asynchrony, communication), and privacy (blind computation, differential privacy). The graph half contributes a taxonomy of graph partitions—edge-cut and vertex-cut topology splits, feature partitions, and heterogeneous-schema partitions—together with graph-specific statistical heterogeneity, adding structure shift. A unified framework culminates in Algorithm 3 for federated training of subgraph GNNs that supports both isolated and collaborative inference, framing subgraph learning as a hybrid of horizontal and vertical FL. Table 2 maps the literature—citation numbers up to 447, implying several hundred surveyed works—onto partition types and techniques, and the authors state that coverage rests on keyword searches for 'federated graph' and 'decentralized graph learning'; resources are on GitHub. Limitations: scope restricted to supervised learning; the effectiveness-efficiency-privacy trilemma is analyzed qualitatively, with no empirical benchmarking of surveyed methods; vertical topology partitions are noted as lacking distributed implementations; malicious adversaries are explicitly out of scope. It reads best as a rigorous research map and entry point rather than a practical tuning guide.
Review complete
Assessment at a glance
3
risk flags identified
● high 0
● medium 1
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- No existing survey systematically analyzes the intersection of collaborative learning and graph-structured data (§1.1, Table 1).
- Collaborative learning on graphs stands on three pillars: network diffusion algorithms, graph representation learning, and collaborative learning (§1.1, Figure 2).
- Design choices across both Euclidean and graph regimes are organized by an effectiveness-efficiency-privacy trilemma (Figure 7, carried through Sections 2.3-2.5 and 3.4-3.6).
- Graph partitions require a richer taxonomy than Euclidean horizontal/vertical splits: edge-cut/vertex-cut topology partitions, feature partitions, and schema partitions (§3.3.1).
- Collaborative learning on subgraphs borrows from horizontal FL when inference is isolated and from vertical FL when inference is collaborative, unified in Algorithm 3 (§3.2).
Strengths
Original unifying taxonomy for an emerging intersection
§3.3.1's graph partition taxonomy (edge-cut/vertex-cut topology, feature, and schema partitions) generalizes the Euclidean horizontal/vertical dichotomy of §2.2.1, and Table 2 operationalizes it by mapping the literature — citation numbers up to 447 — onto partition-technique cells; §1.1 and Table 1 argue no prior survey covers this intersection.
Implementation-grade unified algorithm
Algorithm 3 specifies federated subgraph GNN training with foreign/private boundary node sets (Vf,k, Vp,k), per-layer embedding exchange over the server (lines 11-12), and server-side representation aggregation (line 23), unifying isolated and collaborative inference; §3.2 explicitly derives subgraph learning from horizontal FL (isolated) and vertical FL (collaborative) principles.
Disciplined trade-off structure
Graph-specific heterogeneity characterization
Anchored in theory and deployed practice
Weaknesses
Qualitative-only trilemma analysis
Coverage protocol undocumented
Supervised-only scope
Editorial polish
Acknowledged but undeveloped gap
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This TMLR-published survey (arXiv 2609.02984; cs.LG/cs.MA/cs.SI) consolidates collaborative learning — federated and decentralized learning — into one design space spanning Euclidean data (Section 2) and graph-structured data (Section 3). Its central claims: (1) no prior survey systematically analyzes this intersection (§1.1, Table 1); (2) the field rests on three pillars — network diffusion algorithms, graph representation learning, collaborative learning (Figure 2); (3) design choices are organized by a trilemma of effectiveness, efficiency, and privacy (Figure 7), applied uniformly across both data regimes; (4) graphs require a richer partition taxonomy — edge-cut/vertex-cut topology, feature, and schema partitions (§3.3.1) — plus a new heterogeneity dimension, structure shift (§3.3.2); (5) subgraph collaborative learning unifies horizontal FL principles (isolated inference) and vertical FL principles (collaborative inference), formalized in Algorithm 3. Scope is explicitly limited to supervised learning (§1.1); malicious adversaries are excluded (§3.3.4). The abstract promises systematic identification of open challenges and points to a public GitHub repository.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 3 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the paper 'From Euclidean to Graph-Structured Data: A Survey of Collaborative Learning' about?
It is a TMLR-published survey that organizes federated and decentralized learning research, moving from standard Euclidean data (images, text) to graph-structured data. It reviews methods along three axes — learning effectiveness, efficiency, and privacy — introduces a taxonomy of how graphs can be partitioned across agents (edge-cut/vertex-cut, feature, and schema partitions), and gives a unified algorithm for federated training of graph neural networks on distributed subgraphs.
Does the survey come with code or other resources?
Yes. The paper states that resources are available in a public GitHub repository (remibourgerie/collaborative_gnns), accompanying the taxonomy and the unified training algorithm (Algorithm 3) presented in the survey.
How is this survey different from earlier federated graph learning surveys?
According to the authors (Section 1.1, Table 1), prior surveys cover either federated learning on Euclidean data or, to a lesser extent, graph data, but none systematically map the intersection. This survey connects three research pillars — network diffusion algorithms, graph representation learning, and collaborative learning — and maps the literature (citation numbers up to 447) onto partition types and techniques in a single landscape table.
What open problems does the survey identify?
Per the abstract and Section 3.3, notable gaps include: vertical topology partitions of graphs still lack distributed training implementations; the effectiveness-efficiency-privacy trade-offs are characterized only qualitatively; and the survey restricts itself to supervised learning while excluding malicious-adversary threat models — leaving supervised, honest-but-curious settings as its core coverage with open challenges listed in a dedicated section.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The reviewed text ends mid-§3.4.1 (graph data augmentation); §3.4.2 onward — remaining effectiveness/efficiency/privacy sections, the open-challenges section, the references, and Appendix A — were not visible, so those parts are assessed only via the abstract and introduction.
- Two interior sampling gaps hide portions of §2.3-2.5 (Euclidean effectiveness/efficiency/privacy detail) and §3.1-3.4 (MPNN formalism and graph effectiveness), so judgments there rest on adjacent visible content.
- Table 2 assigns works by numeric citation IDs; the reference list resolving those IDs was not in the visible text, so individual cell assignments could not be audited.
- Some apparent typographical errors (e.g., duplicated words in §2.2.2) could partly stem from PDF text extraction rather than the source.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.