Source-linked AI summary
FederatedScope-GNN: Towards a Unified, Comprehensive and Efficient Package for Federated Graph Learning
Zhen Wang, Weirui Kuang, Yuexiang Xie, Liuyi Yao, Yaliang Li, Bolin Ding, Jingren Zhou
TL;DR
Federated graph learning lacks dedicated framework support despite graph data’s prevalence, leaving benchmark and deployment needs unmet. FS-G packages modular FGL development, benchmarks, tuning, and privacy tools; the authors report collaborative GNN improvements over local learning in three real-world E-commerce scenarios.
Problem
FGL lacks dedicated framework capabilities and comparable benchmarks, while its additional graph-related information exchanges raise privacy-assessment needs.
Source: §1 · pp. 1–2Method
FS-G builds on FederatedScope and combines modular algorithm expression, graph data and model resources, model tuning, and plug-in privacy attacks and defenses.
Source: §1 · p. 2, §3 · p. 3, §3.2 · p. 3Results
In three real-world E-commerce scenarios, collaboratively learned GNNs outperformed their locally learned counterparts.
Source: §1 · p. 2Takeaways & Limitations
Abstract
from arXiv · showhide
The incredible development of federated learning (FL) has benefited various tasks in the domains of computer vision and natural language processing, and the existing frameworks such as TFF and FATE has made the deployment easy in real-world applications. However, federated graph learning (FGL), even though graph data are prevalent, has not been well supported due to its unique characteristics and requirements. The lack of FGL-related framework increases the efforts for accomplishing reproducible research and deploying in real-world applications. Motivated by such strong demand, in this paper, we first discuss the challenges in creating an easy-to-use FGL package and accordingly present our implemented package FederatedScope-GNN (FS-G), which provides (1) a unified view for modularizing and expressing FGL algorithms; (2) comprehensive DataZoo and ModelZoo for out-of-the-box FGL capability; (3) an efficient model auto-tuning component; and (4) off-the-shelf privacy attack and defense abilities. We validate the effectiveness of FS-G by conducting extensive experiments, which simultaneously gains many valuable insights about FGL for the community. Moreover, we employ FS-G to serve the FGL application in real-world E-commerce scenarios, where the attained improvements indicate great potential business benefits. We publicly release FS-G, as submodules of FederatedScope, at https://github.com/alibaba/FederatedScope to promote FGL's research and enable broad applications that would otherwise be infeasible due to the lack of a dedicated package.
1 INTRODUCTION
Federated graph learning lacks the dedicated framework and unified benchmarks available for vision and language, complicating reproducible evaluation and deployment. FS-G addresses this gap with modular algorithm expression, benchmarks, tuning, privacy tools, and reported E-commerce applications.
1 INTRODUCTION
- Existing FL frameworks lack off-the-shelf FGL capabilities and comparable benchmarks, while graph-learning behavior of FL optimizers remains unclear.FGL works also implement methods from scratch on separate testbeds, increasing engineering effort and the risk of unfair comparisons.Source: §1 · pp. 1–2
- FS-G builds on FederatedScope’s event-driven messages and handlers so developers can express heterogeneous exchanges and participant behaviors modularly.Handlers can use graph-learning backends such as torch_geometric and tf_geometric.Source: §1 · p. 2
- GraphDataZoo and GNNModelZoo integrate dataset-splitting mechanisms and state-of-the-art FGL algorithms to support benchmarking and reproduction.The paper also identifies graph-structure-induced covariate shift and designs a federal random graph model for further study.Source: §1 · p. 2
- FS-G provides low-fidelity HPO, client-dissimilarity metrics and parameter grouping for personalization, plus plug-in privacy attacks and defenses.These address costly tuning under non-i.i.d. graph data and assessment of leakage from additional exchanged graph information.Source: §1 · p. 2
- In three real-world E-commerce scenarios, collaboratively learned GNNs outperformed their locally learned counterparts.The authors report that these improvements indicate business value for FS-G.Source: §1 · p. 2
2 RELATED WORK
Related work covers federated optimization, graph-specific learning challenges, and FL software. Existing FL approaches and frameworks offer limited direct support for the heterogeneous exchanges and broader needs of FGL.
2.1 Federated Learning
- FedAvg [28] and later FL optimizers are extensively studied on vision and language, but their behavior when optimizing GNNs is less understood.The surveyed methods include FedProx, FedOPT, FedNOVA, and SCAFFOLD [16].Source: §2.1 · p. 2
2.2 Federated Graph Learning
- FGL methods address cross-client edges [42], heterogeneous graph-level tasks [37], subgraph augmentation [36], and entity alignment [30], often exchanging heterogeneous data.These exchanges entail richer participant behaviors than ordinary FL.Source: §2.2 · pp. 2–3
2.3 FL Software
- Most FL frameworks use declarative computational graphs and lack dedicated FGL support; FedGraphNN [13] covers only simpler canonical behaviors.The surveyed frameworks also omit efficient GNN tuning and FGL privacy attacks and defenses.Source: §2.3 · p. 3
3 INFRASTRUCTURE
FS-G uses FederatedScope’s event-driven framework to modularize FGL communication and participant behavior, and builds graph data and model resources, a unified execution interface, tuning, and privacy components.
3 INFRASTRUCTURE
- FS-G is built on FederatedScope, integrates graph-learning backends and FGL algorithms, and provides a Runner unifying simulation and distributed execution.It also includes model tuning and privacy attack and defense components.Source: §3 · p. 3
3.1 Requirements of Federated Graph Learning
- FGL algorithms exchange heterogeneous data over multiple passes and require richer participant subroutines than general FL’s homogeneous exchanges and canonical updates or aggregation.A unified view can make developers communication-agnostic, modularize procedures, and allow flexible backend choice.Source: §3.1 · p. 3
3.2 Development based on FederatedScope
- FederatedScope represents exchanges as messages and server/client behaviors as handlers; FS-G uses these to implement FGL procedures such as FedSage+ and other methods.FedSage+ messages include model parameters, gradients, and node embeddings; callback functions handle received messages and trigger subroutines.Source: §3.2 · pp. 3–4
4 GRAPHDATAZOO
GraphDataZoo provides configurable components for constructing federated graph datasets from standalone graphs, alongside new datasets and a model for isolating structural covariate shift.
4 GRAPHDATAZOO
- GraphDataZoo combines configurable datasets, splitters, transforms, and dataloaders to support a unified, flexible FGL testbed.Splitters distribute standalone graphs across clients with configurable heterogeneity; node/link tasks use client subgraphs, while graph-level tasks use subsets of graphs.Source: §4 · p. 4, §4.1 · p. 4
4.1 Splitting Standalone Datasets
- Its splitters support metadata- or attribute-based partitions and simulate covariate shift, concept drift, and prior probability shift.The splitters distribute a standalone graph dataset among clients with configurable statistical heterogeneity.Source: §4.1 · p. 4
4.2 New Federated Learning Datasets
- FS-G adds FedDBLP and cross-scenario recommendation datasets, with splits by venue or organizer and by item category or interaction scenario, respectively.FedDBLP represents papers as nodes, citations as edges, abstract bag-of-words as node attributes, and paper themes as labels; the recommendation graph comes from E-commerce user-item interactions.Source: §4.2 · p. 4
- FedcSBM decouples structural from attribute covariate shift by keeping client node-attribute distributions the same while allowing homophily to differ.The federal random graph model is based on cSBM [10].Source: §4.2 · p. 4
5 GNNMODELZOO AND MODEL-TUNING COMPONENT
FS-G combines modular GNN components with FGL-specific tuning, monitoring, and personalization tools. Its HPO support lowers evaluation fidelity, while client metrics and parameter grouping support adaptation to non-i.i.d. graphs.
5 GNNMODELZOO AND MODEL-TUNING COMPONENT
- The GNNModelZoo modularizes models into encoders, GNNs, decoders, and readouts, and includes architectures and algorithms for out-of-the-box FGL.It also includes propagation/feature-transformation-decoupled GNNs such as GPR-GNN [8], which the authors identify as useful for structural covariate shift among client graphs.Source: §5 · pp. 4–5
5.1 Federated Hyper-parameter Optimization
- Low-fidelity HPO reduces FGL trial cost by running fewer communication rounds and sampling a fraction of clients, then retaining the top half of configurations at each SHA stage.Trials can restore a course from a checkpoint, run limited rounds, and evaluate performance; the process continues until one configuration remains.Source: §5.1 · p. 5, Figure 4 · p. 5
- Checkpoints preserve server round and model state, stateful aggregator values, client data-loader state, and personalized client parameters for restoring an FGL course.FS-G exposes a callable runner returning configurable metrics, allowing one-shot HPO methods to be generalized to FGL and system metrics such as average round latency to be optimized.Source: §5.1 · p. 5
5.2 Monitoring and Personalization
- B-local dissimilarity and gradient covariance monitor client heterogeneity; B-local dissimilarity is larger on non-i.i.d. than i.i.d. data at almost all training stages.FS-G logs and visualizes built-in or user-registered metrics through monitoring toolkits.Source: §5.2 · pp. 5–6
- When metrics indicate non-i.i.d. data, users can personalize model parameters or hyperparameters, declaring model parts client-specific or shared so only shared parts are exchanged.The example assigns each client its own encoder and decoder for tasks with different node attributes and classes; finer-grained layer- or variable-level personalization is also possible.Source: §5 · p. 5, §5.2 · p. 6
6 OFF-THE-SHELF ATTACK AND DEFENCE ABILITIES
FS-G integrates passive and malicious privacy attacks with defenses for federated learning. Its GNNTrainer hooks support attacks during training, while plug-in strategies and message checks address passive and malicious attacks.
6 OFF-THE-SHELF ATTACK AND DEFENCE ABILITIES
- Optional GNNTrainer hooks integrate passive class-representative, membership, property, and training-input-and-label inference attacks.The selected hook receives the GNN model and needed target-data information during training.Source: §6 · p. 6
- FS-G also supports malicious adversaries that deviate from the FL protocol by modifying messages.FederatedScope's information-checking mechanism detects anomalous messages to defend against malicious attacks.Source: §6 · p. 6
- For passive privacy attacks, FS-G can use plug-in defenses including differential privacy, MPC, and data compression.Source: §6 · p. 6
7 EXPERIMENTS
The experiments benchmark FGL across node-, link-, and graph-level tasks, study hyperparameter tuning and personalization, and test deployment in three E-commerce scenarios. Across these settings, results vary with the split, task, and degree of client non-I.I.D.ness.
7.1.1 Node-level Tasks.
- FGL generally improves on local training for node-level tasks; global training often leads on random splits, while FGL often leads on community splits.The authors attribute the community-split pattern to removed edges between differently labeled nodes, which increases homophily.Source: §7.1.1 · p. 6, Table 3 · p. 7, Table 2 · p. 7
7.1.2 Link-level Tasks.
- On link-level tasks, FGL outperforms local training and achieves performance comparable to global training.The authors connect broad local/FGL gaps to unbalanced relation distributions that leave clients with insufficient triplets.Source: §7.1.2 · p. 7, Table 4 · p. 7
7.1.3 Graph-level Tasks.
- Graph-level results generally preserve node-level setting relationships, but severe non-I.I.D.ness can leave FedAvg below local training even after personalization.FedOpt and FedProx sometimes surpass local training on the multi-task dataset; label imbalance makes accuracy identical across GNNs on HIV, so ROC-AUC is reported instead.Source: §7.1.3 · p. 7, Table 5 · p. 7
7.2 Study about Hyper-parameter Optimization
- Higher-fidelity SHA searches find better GCN and GPR-GNN configurations, while low-fidelity trials eliminate many poor configurations.Increasing client sampling or training rounds offers comparable gains in search quality, allowing practitioners to trade these costs off.Source: Figure 6 · p. 8, §7.2 · p. 8
7.3 Study about Non-I.I.D.ness and Personalization
- Personalization outperforms ordinary FL and local training across clients with different homophilic degrees.Ordinary FL is comparable to local training; the authors suggest FedAvg may fail to converge because of dissimilar received parameters.Source: §7.3 · p. 8, Figure 7 · p. 8
7.4 Deployment in Real-world E-commerce Scenarios
- In three E-commerce scenarios, federated GraphSAGE reaches ROC-AUC 0.6278±0.0040 versus 0.6209±0.0024 for individually trained models.The authors state that an offline ROC-AUC improvement at the 0.001 level can correspond to a substantial real-world click-through-rate difference.Source: §7.4 · pp. 8–9
8 CONCLUSION
FS-G packages tools for expressing, benchmarking, and tuning federated graph learning algorithms, with privacy utilities and experimental and deployment studies.
8 CONCLUSION
- FS-G provides a unified way to express FGL algorithms, validate them against unified benchmarks, and tune them efficiently.It also includes plug-in attack and defense utilities for assessing privacy leakage.Source: §8 · p. 9
- The paper reports business benefits from deploying FS-G in real-world E-commerce scenarios.Source: §8 · p. 9
- The authors plan to release FS-G to support FGL research and applications.Source: §8 · p. 9
A.1 Details of Off-the-shelf Splitters
FS-G's splitters create federated graph data under different assumptions about client partitions and sources of heterogeneity.
A.1 Details of Off-the-shelf Splitters
- Community, random, and metadata-based splitters partition graphs by detected clusters, random node subsets, or interpretable metadata and attributes.Community splitting can balance client node counts; random splitting can allow intersections or remove a specified fraction of edges.Source: §A.1 · p. 11
- Instance-space, label-space, and multi-task splitters model feature skew, label skew, or clients with distinct tasks and graph domains.Label-space splitting uses LDA for classification triplets and can discretize regression labels before LDA.Source: §A.1 · p. 11
A.2 Details about Our Experiments
The appendix adds experimental settings and results on FGL methods, datasets, and deployment-related tasks.
A.2 Details about Our Experiments
- Node-level experiments use 400 training rounds, two GNN layers, and five clients on citation networks; the HPO study targets PubMed with five clients.The personalization study generates eight FedcSBM graphs at different homophily levels, while the appendix also provides dataset statistics and HIV ROC-AUC reporting.Source: §A.2 · p. 11, Table 6 · p. 12, Table 10 · p. 12
- FedSage+ significantly outperforms its GraphSAGE baseline on most reported node-classification datasets.The method jointly trains generative models to predict missing links, enabling clients to reconstruct missing cross-client links.Source: §A.2 · p. 11, Table 7 · p. 12
- On FedDBLP, FGL achieves performance comparable to global training despite small client graphs, removed edges, and more clients than in prior experiments.The authors present this as evidence that FGL can handle simulated data interruption.Source: §A.2 · p. 11, Table 8 · p. 12
- GCFL+ outperforms FedAvg on IMDB and Multi-task datasets and is comparable on PROTEINS.GCFL+ groups clients by gradient sequences so clients in a cluster share model parameters; the authors suggest this may handle client non-I.I.D.ness better.Source: §A.2 · p. 11
A.3 Datasets description
FS-G describes its datasets alongside suggested Splitters, noting that datasets span domains and that node and edge meanings vary. The package plans to add datasets and benchmarks.
A.3 Datasets description
- FS-G pairs its dataset descriptions with suggested Splitters.The passage also references Table 6.Source: §A.3 · p. 12
- Table 11 compares GCFL+ and GIN (with FedAvg) on graph classification datasets using mean accuracy (%) ± standard deviation.Source: Table 11 · p. 12
- The datasets span different domains, and their nodes and edges represent different meanings.Source: §A.3 · p. 12
- The package plans to support more datasets and provide more benchmarks in the future.Source: §A.3 · p. 12