FastPFRec: A Fast Personalized Federated Recommendation with Secure Sharing

Zhenxing Yan; Jidong Yuan; Yongqi Sun; Haiyang Liu; Zhihui Gao

cs.IRMar 18, 2026

FastPFRec: A Fast Personalized Federated Recommendation with Secure Sharing

Authors: Zhenxing Yan, Jidong Yuan, Yongqi Sun, Haiyang Liu, Zhihui Gao

Organizations: Key Laboratory of Big Data & Artificial Intelligence in Transportation (Beijing Jiaotong University), Ministry of Education, Beijing, 100044, China · School of Computer Science and Technology, Beijing Jiaotong University, Beijing, 100044, China

Abstract

Graph neural network (GNN)-based federated recommendation systems effectively capture user-item relationships while preserving data privacy. However, existing methods often face slow convergence on graph data and privacy leakage risks during collaboration. To address these challenges, we propose FastPFRec (Fast Personalized Federated Recommendation with Secure Sharing), a novel framework that enhances both training efficiency and data security. FastPFRec accelerates model convergence through an efficient local update strategy and introduces a privacy-aware parameter sharing mechanism to mitigate leakage risks. Experiments on four real-world datasets (Yelp, Kindle, Gowalla-100k, and Gowalla-1m) show that FastPFRec achieves 32.0% fewer training rounds, 34.1% shorter training time, and 8.1% higher accuracy compared with existing baselines. These results demonstrate that FastPFRec provides an efficient and privacy-preserving solution for scalable federated recommendation.

Figures & tables

Figure 1: An example of federated attack with noise injection.

Figure 2: Comparison of training NDCG across epochs on the Gowalla-1m dataset.

Figure 3: The proposed FastPFRec framework. This flowchart illustrates the end-to-end process, highlighting the local training with FastGNN on client devices and the subsequent aggregation of model updates by trusted nodes before being sent to the central server.

Figure 4: The position of trusted node in the overall framework.

Dataset	# User	# Item	# Interaction	Density
Yelp	5,145	7,421	124,036	0.32%
Kindle	7,650	9,173	137,124	0.20%
Gowalla-100k (Subset)	3,917	7,000	85,442	0.31%
Gowalla-1m (Subset)	29,858	40,981	810,128	0.04%

Table 1: Statistics of datasets. Note that Gowalla-100k and Gowalla-1m are dense subsets derived from the original Gowalla dataset to simulate different scales.

Centralized Models
Model	Yelp		Kindle		Gowalla-100k		Gowalla-1m
Model	HR	NDCG	HR	NDCG	HR	NDCG	HR	NDCG
MF	$0.0706_{\pm 0.0015}$	$0.0286_{\pm 0.0012}$	$0.1519_{\pm 0.0023}$	$0.0957_{\pm 0.0018}$	$0.1664_{\pm 0.0021}$	$0.1461_{\pm 0.0019}$	$0.1021_{\pm 0.0018}$	$0.1085_{\pm 0.0016}$
LightGCN	$0.0860_{\pm 0.0018}$	$0.0339_{\pm 0.0014}$	$0.1874_{\pm 0.0025}$	$0.1177_{\pm 0.0020}$	$0.1842_{\pm 0.0023}$	$0.1658_{\pm 0.0021}$	$0.1303_{\pm 0.0020}$	$0.1435_{\pm 0.0018}$
SGL	$0.0911_{\pm 0.0019}$	$0.0349_{\pm 0.0015}$	$0.1886_{\pm 0.0026}$	$0.1188_{\pm 0.0021}$	$0.1867_{\pm 0.0024}$	$0.1692_{\pm 0.0022}$	$0.1364_{\pm 0.0021}$	$0.1471_{\pm 0.0019}$
XSimGCL	$0.0959_{\pm 0.0020}$	$0.0369_{\pm 0.0016}$	$0.1767_{\pm 0.0024}$	$0.1120_{\pm 0.0019}$	$0.1896_{\pm 0.0025}$	$0.1750_{\pm 0.0023}$	$0.1466_{\pm 0.0022}$	$0.1576_{\pm 0.0020}$

Table 2: Performance comparison of different models on Yelp, Kindle, and Gowalla datasets (including Gowalla-100k and Gowalla-1m). Results are reported as mean

\pm

standard deviation across 5 runs. “*” indicates statistical significance at

p<0.05

in Wilcoxon signed-rank test against baseline methods.

Figure 5: CD diagram on HR.

Structural Modules
Model / Variant	Yelp		Kindle		Gowalla-1m
Model / Variant	HR@10	NDCG@10	HR@10	NDCG@10	HR@10	NDCG@10
FastPFRec (Default)	0.0705	0.0271	0.1399	0.0861	0.0997	0.1073
w/o FastGNN (use LightGCN)	0.0642	0.0242	0.1331	0.0782	0.0912	0.0923
w/o Scheduled Item Updating (Full Update)	0.0648	0.0246	0.1340	0.0790	0.0925	0.0938
Security & Privacy Modules

Table 3: Expanded Ablation Study: Performance of FastPFRec and its variants. "w/o" denotes the removal of a specific module. Note that "w/o LDP" represents the non-private baseline.

Panel of Figure 6 — Figure 6: Training loss of FastPFRec and PerFedRec++ on Yelp and Gowalla-1m.

Panel of Figure 7 — Figure 7: Training NDCG of FastPFRec and PerFedRec++ on Yelp and Gowalla-1m (per epoch).

Figure 8: Comparison of epochs for federated recommendation models achieving the best NDCG across different datasets.

Privacy Parameters		Utility Metric
$P_{\text{pert}}$	$\lambda$ (LDP Noise)	NDCG	Drop Rate (%)
0.10	0 (No Noise)	0.1120	-
0.10	0.05	0.1108	1.07%
0.10	0.10 (Default)	0.1073	4.29%
0.10	0.20	0.1065	4.91%

Table 4: Privacy-Utility Trade-off on Gowalla-1m. We fix the graph perturbation

P_{\text{pert}}=0.1

and vary the LDP noise scale

\lambda

. The row with

\lambda=0.10

corresponds to our default setting.

Architecture	Server Damage Avg	Protection Rate
Traditional (Direct)	$2.34_{\pm 0.15}$	0.0%
FastPFRec (w/ Trusted Nodes)	$0.12_{\pm 0.08}$	94.8%

Table 5: Attack resilience comparison: Traditional vs. FastPFRec architecture

Failed Nodes	Failure Rate	NDCG
0	0%	$0.1073_{\pm 0.0002}$
1	10%	$0.1068_{\pm 0.0004}$
2	20%	$0.1070_{\pm 0.0008}$
3	30%	$0.1064_{\pm 0.0005}$

Table 6: System performance under trusted node failures

Compromised Nodes	NDCG	Containment Rate
0	$0.1073_{\pm 0.0013}$	100.0%
1	$0.1069_{\pm 0.0009}$	99.2%
2	$0.1065_{\pm 0.0010}$	98.7%
3	$0.1062_{\pm 0.0008}$	98.1%

Table 7: Attack resilience under node compromise

Attack Type	Detection Rate	False Positive	Latency (ms)
Noise Injection	97.3%	3.2%	$38\pm 5$
Gradient Poisoning	96.8%	4.1%	$46\pm 6$

Table 8: Anomaly detection performance metrics

Model	Time Complexity
FedMF	$\mathcal{O}(\|U\|\cdot\|I\|\cdot k)$
FedNCF	$\mathcal{O}(H\cdot\|U\|\cdot\|I\|\cdot k)$
FedGNN	$\mathcal{O}(H\cdot\|U\|\cdot\|I\|\cdot k)$
PerFedRec	$\mathcal{O}(H\cdot\|U\|\cdot\|I\|\cdot k)$
PerFedRec++	$\mathcal{O}(H\cdot\|U\|\cdot\|I\|\cdot k)$
FastPFRec	$\mathcal{O}(H\cdot k\cdot(\|U\|+\|I\|/h))$

Table 9: Time complexity of different federated client models (per epoch).

Figure 9: Time consumption of federated recommendation models over same epochs.

Panel of Figure 10 — Figure 10: Performance comparison with different embedding size.

$T$	Yelp			Kindle			Gowalla-1m
$T$	HR	NDCG	$\mathcal{T}_{500}$	HR	NDCG	$\mathcal{T}_{500}$	HR	NDCG	$\mathcal{T}_{500}$
5	0.0688	0.0261	3303	0.1374	0.0851	3194	0.0973	0.1032	21933
10	0.0704	0.0270	3429	0.1398	0.0860	3204	0.0996	0.1073	22563
20	0.0682	0.0268	3518	0.1361	0.0858	3317	0.0972	0.1029	24016

Table 10: Performance and time cost (in seconds) comparison with different numbers of trusted nodes.

Model	Yelp	Kindle	Gowalla-1m
FastPFRec	9.35	8.23	12.09
PerFedRec++	9.59	7.87	13.24

Table 11: Communication time in real-world simulation (in seconds)

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 502 papers

Sharpness-Aware Minimization for Generalized Embedding Learning in Federated Recommendation
Mar 12, 2026Fengyuan Yu, Xiaohua Feng, Yuyuan Li +3Recommender Systems Non-IID Federated Learning
Zhejiang University, Hangzhou, China · Hangzhou Dianzi University, Hangzhou, China · OPPO Research Institute, Shenzhen, China
Provably Communication-Efficient and Privacy-Preserving Federated Graph Neural Networks
May 25, 2026Zhishuai Guo, Wenhan Wu, Chen Chen +3Graph Neural Networks Communication-Efficient Distributed Training
Department of Computer Science Northern Illinois University · Department of Computer Science University of North Carolina at Charlotte · Department of Computer Science University of Central Florida +2
Low-pass Personalized Subgraph Federated Recommendation
Mar 20, 2026Wooseok Sim, Hogun ParkPopularity Bias in Recommender Systems Recommender Systems
Department of Artificial Intelligence, Sungkyunkwan University, Suwon, Republic of Korea

Mar 12, 2026J/K move · Enter open · S save

Sharpness-Aware Minimization for Generalized Embedding Learning in Federated Recommendation

Fengyuan Yu, Xiaohua Feng, Yuyuan Li, Changwang Zhang+2

Zhejiang University, Hangzhou, China · Hangzhou Dianzi University, Hangzhou, China · OPPO Research Institute, Shenzhen, China

Recommender Systems Non-IID Federated Learning Heterogeneous Federated Learning

Open paper AskPDF arXiv

Federated recommender systems enable collaborative model training while keeping user interaction data local and sharing only essential model parameters, thereby mitigating privacy risks. However, existing methods overlook a critical issue, i.e., the stable learning of a generalized item embedding throughout the federated recommender system training process. Item embedding plays a central role in facilitating knowledge sharing across clients. Yet, under the cross-device setting, local data distributions exhibit significant heterogeneity and sparsity, exacerbating the difficulty of learning generalized embeddings. These factors make the stable learning of generalized item embeddings both indispensable for effective federated recommendation and inherently difficult to achieve. To fill this gap, we propose a new federated recommendation framework, named Federated Recommendation with Generalized Embedding Learning (FedRecGEL). We reformulate the federated recommendation problem from an item-centered perspective and cast it as a multi-task learning problem, aiming to learn generalized embeddings throughout the training procedure. Based on theoretical analysis, we employ sharpness-aware minimization to address the generalization problem, thereby stabilizing the training process and enhancing recommendation performance. Extensive experiments on four datasets demonstrate the effectiveness of FedRecGEL in significantly improving federated recommendation performance. Our code is available at https://github.com/anonymifish/FedRecGEL.

Abstract

Recommender Systems

Explore similar work

Sharpness-Aware Minimization for Generalized Embedding Learning in Federated Recommendation

Provably Communication-Efficient and Privacy-Preserving Federated Graph Neural Networks

Low-pass Personalized Subgraph Federated Recommendation