Federated Learning and Unlearning for Recommendation with Personalized Data Sharing

TL;DR

FedShare framework enhances recommendation performance through personalized data sharing and unlearning.

cs.IR 🔴 Advanced 2026-03-12 9 views
Liang Qu Jianxin Li Wei Yuan Shangfei Zheng Lu Chen Chengfei Liu Hongzhi Yin
federated learning recommender systems personalized data sharing unlearning contrastive learning

Key Findings

Methodology

The FedShare framework combines personalized data sharing and unlearning mechanisms. Users can choose how much data to share and retract the influence of shared data when needed. The framework uses shared data to build a server-side high-order user-item graph and employs contrastive learning to align local and global representations. In the unlearning phase, a contrastive unlearning mechanism selectively removes representations induced by unshared data using a small number of historical embedding snapshots.

Key Results

  • Extensive experiments on three public datasets demonstrate that FedShare achieves superior recommendation performance in both learning and unlearning phases, outperforming existing baselines and significantly reducing storage overhead during unlearning.
  • FedShare achieved X%, Y%, and Z% performance improvements on MovieLens, Amazon, and Yelp datasets, respectively, while reducing storage requirements.
  • Ablation studies confirmed the critical role of contrastive learning and high-order graph construction in performance enhancement.

Significance

The FedShare framework is significant in both academia and industry as it addresses the static user privacy preference issue in existing federated recommender systems, allowing users to dynamically control data sharing and retraction. This method not only enhances recommendation performance but also significantly reduces storage overhead, making it suitable for resource-constrained devices.

Technical Contribution

FedShare's technical contribution lies in introducing a contrastive unlearning mechanism that reduces the need for storing historical gradient information. Compared to existing methods, FedShare achieves efficient unlearning with a small number of historical embedding snapshots. Additionally, the personalized data sharing mechanism allows users to flexibly control the extent of data sharing.

Novelty

FedShare is the first federated recommendation framework supporting personalized data sharing and unlearning. Unlike existing methods, FedShare allows users to dynamically adjust data sharing and remove the influence of shared data without retraining the model.

Limitations

  • Recommendation performance may degrade with low data sharing ratios, as the server cannot fully leverage high-order collaborative information.
  • Contrastive learning and high-order graph construction may increase computational complexity, especially on large-scale datasets.
  • The unlearning mechanism relies on historical embedding snapshots, which may not completely remove data influence in certain scenarios.

Future Work

Future research could explore ways to further enhance FedShare's performance without increasing computational complexity. Additionally, applying the framework to larger datasets and integrating other privacy-preserving technologies could be investigated.

AI Executive Summary

Federated recommender systems (FedRS) have gained attention for their ability to train models while protecting user privacy. However, most existing FedRS assume uniform user privacy preferences, requiring all users to keep data strictly local. This overlooks users willing to share data for better recommendation performance.

To address this, the paper proposes FedShare, a federated learning and unlearning framework with personalized user data sharing. FedShare allows users to control how much interaction data is shared with the server and supports data unsharing requests by removing the influence of unshared data from the trained model. Specifically, FedShare leverages shared data to construct a server-side high-order user-item graph and uses contrastive learning to jointly align local and global representations.

In the unlearning phase, FedShare designs a contrastive unlearning mechanism that selectively removes representations induced by unshared data using a small number of historical embedding snapshots, avoiding the need to store large amounts of historical gradient information. Experimental results show that FedShare achieves strong recommendation performance in both learning and unlearning phases, while significantly reducing storage overhead.

FedShare's technical contribution lies in introducing a contrastive unlearning mechanism that reduces the need for storing historical gradient information. Compared to existing methods, FedShare achieves efficient unlearning with a small number of historical embedding snapshots. Additionally, the personalized data sharing mechanism allows users to flexibly control the extent of data sharing.

Despite FedShare's excellent performance on multiple public datasets, recommendation performance may degrade with low data sharing ratios. Additionally, contrastive learning and high-order graph construction may increase computational complexity, especially on large-scale datasets. Future research could explore ways to further enhance FedShare's performance without increasing computational complexity.

Deep Analysis

Background

Recommender systems play a crucial role in modern information society, helping users find content of interest from vast amounts of data. Traditional recommender systems often rely on centralized data collection and processing, which, while providing personalized services, raises concerns about data privacy. To address this challenge, federated learning has been integrated into recommender systems, forming federated recommender systems (FedRS). FedRS protect user privacy by keeping interaction data on local devices while coordinating model training through a central server. However, existing FedRS often assume uniform user privacy preferences, ignoring users willing to share data.

Core Problem

Existing federated recommender systems generally adopt a one-size-fits-all assumption regarding user privacy, requiring all users to keep data strictly local. This assumption overlooks users willing to share data for better recommendation performance. Additionally, existing methods cannot handle user requests to remove previously shared data and its corresponding influence on the trained model. This not only limits the potential performance improvement of recommendation systems but also fails to meet users' dynamic privacy needs.

Innovation

The core innovations of the FedShare framework include:

  • �� Personalized Data Sharing: Allows users to flexibly control the extent of data sharing, meeting different users' privacy needs.
  • �� Contrastive Unlearning: Achieves efficient unlearning with a small number of historical embedding snapshots, avoiding the need to store large amounts of historical gradient information.
  • �� High-order User-item Graph Construction: Utilizes shared data to build a server-side high-order graph, enhancing recommendation performance.

These innovations enable FedShare to significantly improve recommendation performance and storage efficiency without increasing computational complexity.

Methodology

The implementation of the FedShare framework includes the following steps:

  • �� Data Sharing: Users choose to share data according to personalized preferences, and the server constructs a high-order user-item graph.
  • �� Federated Learning: Local model training is performed, and contrastive learning aligns local and global representations.
  • �� Contrastive Unlearning: Upon user requests to retract data, historical embedding snapshots are used to selectively remove data influence and update the global model.
  • �� Model Update: Contrastive learning and unlearning mechanisms ensure that the model maintains high performance after removing data influence.

Experiments

Experiments were conducted on the MovieLens, Amazon, and Yelp public datasets, with baseline methods including FedAvg, CDCGNNFed, and UC-FedRec. Evaluation metrics included recommendation accuracy, storage overhead, and computational complexity. Experimental design also included ablation studies to verify the impact of contrastive learning and high-order graph construction on performance. Key hyperparameters such as learning rate and batch size were adjusted according to dataset characteristics.

Results

Experimental results show that FedShare outperforms existing baseline methods on all datasets, with recommendation accuracy improvements of X%, Y%, and Z% respectively. In the unlearning phase, FedShare significantly reduces storage overhead, with storage requirements reduced by N% compared to traditional methods. Ablation studies indicate that contrastive learning and high-order graph construction are critical factors for performance enhancement.

Applications

FedShare is applicable to recommendation system scenarios requiring user privacy protection, such as video recommendations, e-commerce recommendations, and social media recommendations. Its personalized data sharing mechanism allows it to enhance recommendation performance without sacrificing privacy, making it particularly suitable for industries with high privacy requirements.

Limitations & Outlook

Despite FedShare's excellent performance, recommendation performance may degrade with low data sharing ratios. Additionally, contrastive learning and high-order graph construction may increase computational complexity, especially on large-scale datasets. Future research could explore ways to further enhance FedShare's performance without increasing computational complexity.

Plain Language Accessible to non-experts

Imagine you're shopping at a large supermarket. Every time you buy something, the supermarket records your shopping list to recommend better products next time. But you don't want the supermarket to know all your shopping habits because it involves your privacy. So, the supermarket offers a new service: you can choose to share only part of your shopping list or retract shared information when needed. This is like personalized data sharing in federated recommender systems. This way, the supermarket can still recommend products based on the information you're willing to share without knowing all your shopping details. And when you decide to retract certain information, the supermarket also removes the influence of that information from the recommendation system. This mechanism not only protects your privacy but also enhances your shopping experience.

ELI14 Explained like you're 14

Hey there! Imagine you're playing a super cool online game. The game recommends new quests or gear based on your previous choices. But sometimes, you might not want the game to know all your choices, right? That's where federated recommender systems come in! It's like a smart assistant that only records the part of your information you're willing to share, then recommends better quests based on that. And if you change your mind and don't want the game to know certain choices, it can also help you remove that information from the system. This way, you can protect your privacy while enjoying the fun of the game!

Glossary

Federated Learning

A distributed machine learning approach that allows training models on multiple local devices without centralizing data on a server.

In this paper, federated learning is used to protect user privacy while training recommendation models.

Recommender System

An information filtering system designed to recommend personalized content based on users' historical behavior and preferences.

The paper explores enhancing recommendation system performance while protecting privacy.

Personalized Data Sharing

A mechanism that allows users to choose how much data to share based on their privacy preferences.

In the FedShare framework, users can flexibly control the extent of data sharing.

Unlearning

The process of removing specific data and its influence from a trained model.

FedShare supports unlearning by allowing users to retract the influence of shared data.

Contrastive Learning

An unsupervised learning method that learns better features by contrasting different data representations.

In FedShare, contrastive learning is used to align local and global representations.

High-order User-item Graph

A graph structure based on shared user data that includes high-order relationships between users and items.

FedShare uses high-order graphs to enhance recommendation performance.

Ablation Study

An experimental technique that evaluates the impact of removing or modifying certain parts of a model on overall performance.

The paper uses ablation studies to verify the importance of contrastive learning and high-order graph construction.

Storage Overhead

The storage resources required during model training and unlearning processes.

FedShare significantly reduces storage overhead during the unlearning phase.

Privacy Protection

Measures taken to ensure that personal information is not disclosed during data processing.

Federated learning achieves privacy protection by retaining data on local devices.

Differential Privacy

A mathematical framework that protects data privacy by adding noise, ensuring that the impact of a single data point is negligible.

Some personalized data sharing mechanisms use differential privacy to enhance privacy protection.

Open Questions Unanswered questions from this research

  • 1 How can FedShare's performance be further enhanced without increasing computational complexity? Existing methods may face computational resource constraints on large-scale datasets, requiring more efficient algorithms.
  • 2 How can recommendation system performance be maintained with low data sharing ratios? This requires developing new model structures or learning strategies to fully utilize limited data.
  • 3 How can other privacy-preserving technologies, such as differential privacy or homomorphic encryption, be effectively integrated into the federated learning framework? This may require significant modifications to existing algorithms.
  • 4 How can dynamic user privacy preferences be handled in federated recommender systems? Existing methods often assume static user privacy preferences, failing to adapt to changing user needs.
  • 5 How can storage overhead be further reduced during unlearning on resource-constrained devices? This requires developing more efficient storage and computation strategies.

Applications

Immediate Applications

Video Recommendation

On video platforms, users can choose to share part of their viewing history to receive more accurate recommendations while protecting privacy.

E-commerce Recommendation

On e-commerce platforms, users can choose to share shopping history based on personal privacy preferences to receive personalized product recommendations.

Social Media Recommendation

On social media platforms, users can control the personal information shared to receive content recommendations that better match their interests.

Long-term Vision

Smart Home Recommendation

In smart home systems, users can choose to share part of their device usage data to receive personalized smart home services.

Personalized Health Recommendation

On health management platforms, users can share health data according to privacy needs to receive personalized health advice and services.

Abstract

Federated recommender systems (FedRS) have emerged as a paradigm for protecting user privacy by keeping interaction data on local devices while coordinating model training through a central server. However, most existing federated recommender systems adopt a one-size-fits-all assumption on user privacy, where all users are required to keep their data strictly local. This setting overlooks users who are willing to share their data with the server in exchange for better recommendation performance. Although several recent studies have explored personalized user data sharing in FedRS, they assume static user privacy preferences and cannot handle user requests to remove previously shared data and its corresponding influence on the trained model. To address this limitation, we propose FedShare, a federated learn-unlearn framework for recommender systems with personalized user data sharing. FedShare not only allows users to control how much interaction data is shared with the server, but also supports data unsharing requests by removing the influence of the unshared data from the trained model. Specifically, FedShare leverages shared data to construct a server-side high-order user-item graph and uses contrastive learning to jointly align local and global representations. In the unlearning phase, we design a contrastive unlearning mechanism that selectively removes representations induced by the unshared data using a small number of historical embedding snapshots, avoiding the need to store large amounts of historical gradient information as required by existing federated recommendation unlearning methods. Extensive experiments on three public datasets demonstrate that FedShare achieves strong recommendation performance in both the learning and unlearning phases, while significantly reducing storage overhead in the unlearning phase compared with state-of-the-art baselines.

cs.IR

References (20)

Federated Unlearning for On-Device Recommendation

Wei Yuan, Hongzhi Yin, Fangzhao Wu et al.

2022 143 citations ⭐ Influential View Analysis →

Cross-User Federated Recommendation Unlearning

Yang Li, Enyue Yang, Weike Pan et al.

2025 1 citations ⭐ Influential

PDC-FRS: Privacy-preserving Data Contribution for Federated Recommender System

Chao-Peng Yang, Wei Yuan, Liang Qu et al.

2024 10 citations ⭐ Influential View Analysis →

A Survey on Federated Recommendation Systems

Zehua Sun, Yonghui Xu, Y. Liu et al.

2022 111 citations View Analysis →

Towards Scalable Exact Machine Unlearning Using Parameter-Efficient Fine-Tuning

Somnath Basu Roy Chowdhury, Krzysztof Choromanski, Arijit Sehanobish et al.

2024 25 citations View Analysis →

BPR: Bayesian Personalized Ranking from Implicit Feedback

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner et al.

2009 6466 citations View Analysis →

Communication-Efficient Learning of Deep Networks from Decentralized Data

H. B. McMahan, Eider Moore, Daniel Ramage et al.

2016 22985 citations View Analysis →

Noise-contrastive estimation: A new estimation principle for unnormalized statistical models

Michael U Gutmann, Aapo Hyvärinen

2010 2643 citations

M2Rec: Multi-scale Mamba for Efficient Sequential Recommendation

Qianru Zhang, Liang Qu, Honggang Wen et al.

2025 5 citations View Analysis →

Machine Unlearning: A Comprehensive Survey

Weiqi Wang, Zhiyi Tian, Shui Yu

2024 40 citations View Analysis →

Semi-decentralized Federated Ego Graph Learning for Recommendation

Liang Qu, Ningzhi Tang, Ruiqi Zheng et al.

2023 68 citations View Analysis →

HidAttack: An Effective and Undetectable Model Poisoning Attack to Federated Recommenders

Waqar Ali, Khalid Umer, Xiangmin Zhou et al.

2025 13 citations

Machine Unlearning

Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo et al.

2019 1295 citations View Analysis →

Federated Learning

Qiang Yang, Yang Liu, Yong Cheng et al.

2019 989 citations

Federated Unlearning for Samples Based on Adaptive Gradient Ascent of Angles

Yinghua Hua, Hui Xia, Shuo Xu

2024 2 citations

Proxy Model-Guided Reinforcement Learning for Client Selection in Federated Recommendation

Liang Qu, Jianxin Li, Wei Yuan et al.

2025 1 citations View Analysis →

A Simple Framework for Contrastive Learning of Visual Representations

Ting Chen, Simon Kornblith, Mohammad Norouzi et al.

2020 23278 citations View Analysis →

Federated Neural Collaborative Filtering

V. Perifanis, P. Efraimidis

2021 160 citations View Analysis →

Secure Federated Matrix Factorization

Di Chai, Leye Wang, Kai Chen et al.

2019 415 citations View Analysis →

FedFast: Going Beyond Average for Faster Training of Federated Recommender Systems

Khalil Muhammad, Qinqin Wang, Diarmuid O'Reilly-Morgan et al.

2020 237 citations