FedGNN (Federated Graph Neural Network for Privacy-Preserving Recommendation)
Algorithms AIMech:FedGnnFederatedGraphNeuralNetworkForPrivacyPreservingRecommendation DRAFT
Aliases: FedGNN, Federated Graph Neural Network for Privacy-Preserving Recommendation
FedGNN (Federated Graph Neural Network for Privacy-Preserving Recommendation) is a federated collaborative filtering and graph neural network training framework in which each user client maintains a private first-order user-item interaction subgraph G_i, protects interacted item IDs against server gradient-inversion attacks by sampling M non-interacted pseudo items with synthetic Gaussian embedding gradients g_i^p matching the mean and covariance of real item embedding gradients g_i^e, perturbs the combined model and item-embedding gradients with clipped Laplace local differential privacy g_i = clip(g_i, delta) + Laplace(0, lambda), and expands local graphs to high-order user-item neighbor neighborhoods via homomorphic-encryption item-ID matching and anonymous neighbor user-embedding exchange through a third-party server.
Identity Notes & Concept Scope
Introduced by Chuhan Wu, Fangzhao Wu, Yang Cao, Yongfeng Huang, and Xing Xie in 'FedGNN: Federated Graph Neural Network for Privacy-Preserving Recommendation' (arXiv:2102.04925, DBLP:journals/corr/abs-2102-04925). Distinct from FCF (Federated Collaborative Filtering, Ammad-ud-din et al. 2019), which models only first-order matrix factorization interactions and exposes interacted item IDs via non-zero item-embedding gradients; and distinct from FedSage+ (Zhang et al. NeurIPS 2021), which trains a local missing-neighbor generator (NeighGen) across multi-node subgraphs rather than expanding user-centric bipartite recommendation graphs via encrypted item matching and pseudo-item LDP gradients.
Claims & Grounded Evidence
3 Claims
Mechanism
#fedgnn-pseudo-interacted-item-sampling-and-ldp-gradient-perturbation
To prevent the central server from inferring a user's private interaction history from which item embeddings have non-zero gradients g_i^e and from reconstructing raw ratings y_{i,j} via gradient inversion, each client in FedGNN samples M non-interacted items, generates synthetic pseudo-item gradients g_i^p from a Gaussian distribution matching the empirical mean and covariance of real item-embedding gradients g_i^e (yielding expected interaction anonymity proportional to 1 + M P / |Y_o|), and perturbs the combined gradient g_i = (g_i^m, g_i^e, g_i^p) via g_i = clip(g_i, delta) + Laplace(0, lambda) (Equation 1) with local differential privacy upper bound epsilon <= 2 delta / lambda.
Mechanism
#fedgnn-privacy-preserving-graph-expansion-via-encrypted-item-matching
To capture high-order collaborative user-item signals without revealing user identities or raw interaction lists to the recommendation server or other clients, FedGNN uses Algorithm 2: after T warmup epochs, the central server generates a public/private key pair (p, s) and sends p to clients; each client encrypts its interacted item IDs with p and sends the ciphertexts together with its inferred user embedding e_i^u to a third-party server, which links users sharing identical encrypted item ciphertexts and returns anonymous 1-hop neighbor user embeddings [e_{i,1}^u, ..., e_{i,N}^u] so each client can connect neighbor user nodes to co-interacted items in its expanded local graph G_i.
Application
#fedgnn-recommendation-accuracy-across-six-benchmarks
Evaluated across six benchmark recommendation datasets (MovieLens-100K, MovieLens-1M, MovieLens-10M, Flixster, Douban, and YahooMusic), FedGNN with gated graph neural networks (GGNN), graph convolutional networks (GCN), or graph attention networks (GAT) matches centralized GNN recommendation RMSE and significantly outperforms privacy-preserving Federated Matrix Factorization (FedMF) and Federated Collaborative Filtering (FCF) (Tables 2 and 3, e.g., reducing RMSE on ML-100K from 0.947 for FedMF to 0.910 for FedGNN).
⚠️ Curated Limitations & Negative Results
- Privacy-preserving user-item graph expansion relies on the security assumption that the central recommendation server and the third-party matching server do not collude.
- Uploading pseudo-interacted item gradients g_i^p for M non-interacted items per user increases client-to-server communication cost linearly in M (Figure 6b).
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:FederatedLearning
Federated Learning
|
broad_match | FedGNN is a federated graph neural network training framework for privacy-preserving collaborative recommendation under AIO:FederatedLearning. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r517-primary-pdf
|
Chuhan Wu, Fangzhao Wu, Yang Cao, Yongfeng Huang, Xing Xie
|
primary paper
full_text
|
ID: https://dblp.org/rec/journals/corr/abs-2102-04925
sha256:a9cfe057f622…
|
r517-arxiv-abs
|
Chuhan Wu, Fangzhao Wu, Yang Cao, Yongfeng Huang, Xing Xie
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/2102.04925
sha256:e77ed0ef2b95…
|
r517-metadata-cr
|
Chuhan Wu, Fangzhao Wu, Yang Cao, Yongfeng Huang, Xing Xie
|
primary paper
abstract_only
|
ID: https://dblp.org/rec/journals/corr/abs-2102-04925
sha256:fbeb4afdd158…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T10:00:00Z
Updated: 2026-10-19T10:00:00Z
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories