[Paper Review] Federated Collaborative Filtering for Privacy-Preserving Personalized Recommendation System

Muhammad Ammad-ud-din, Elena Ivannikova+5 more|arXiv (Cornell University)|Jan 29, 2019

Privacy-Preserving Technologies in Data33 references215 citations

TL;DR

The paper introduces the first Federated Collaborative Filter (FCF) that updates item factors on a server and user factors on clients, achieving comparable accuracy to standard CF while preserving user privacy.

ABSTRACT

The increasing interest in user privacy is leading to new privacy preserving machine learning paradigms. In the Federated Learning paradigm, a master machine learning model is distributed to user clients, the clients use their locally stored data and model for both inference and calculating model updates. The model updates are sent back and aggregated on the server to update the master model then redistributed to the clients. In this paradigm, the user data never leaves the client, greatly enhancing the user' privacy, in contrast to the traditional paradigm of collecting, storing and processing user data on a backend server beyond the user's control. In this paper we introduce, as far as we are aware, the first federated implementation of a Collaborative Filter. The federated updates to the model are based on a stochastic gradient approach. As a classical case study in machine learning, we explore a personalized recommendation system based on users' implicit feedback and demonstrate the method's applicability to both the MovieLens and an in-house dataset. Empirical validation confirms a collaborative filter can be federated without a loss of accuracy compared to a standard implementation, hence enhancing the user's privacy in a widely used recommender application while maintaining recommender performance.

Motivation & Objective

• Motivate privacy-preserving personalized recommendations under GDPR-like regimes.
• Develop a federated version of matrix factorization for implicit feedback in recommender systems.
• Demonstrate that federated updates can match standard CF accuracy while keeping user data on-device.
• Provide a privacy-by-design solution without requiring client identities in aggregation.

Proposed method

• Formulate CF with implicit feedback using X (user factors) and Y (item factors) and a confidence matrix C^u.
• Distribute item factors Y on the server and update user factors X locally on each client using local data.
• Compute and aggregate client gradients f(u,i) to update Y on the server via SGD/Adam optimization.
• Use an adaptive Adam-based gradient descent to update Y and ensure convergence to CF optimum.
• Ensure privacy by not requiring user identities and by keeping interactions on-device while exchanging gradients.
• Demonstrate convergence and compare with standard CF on MovieLens, simulated, and in-house datasets.

Experimental results

Research questions

• RQ1 Can a CF model be federated without exposing user interaction data?
• RQ2 Does Federated Collaborative Filtering (FCF) preserve CF accuracy compared to non-federated CF?
• RQ3 What training dynamics and hyperparameters (e.g., learning rate, Adam parameters) ensure stable convergence in FCF?
• RQ4 How does FCF perform on implicit feedback datasets across real and synthetic data?
• RQ5 Is the privacy-preserving design robust to typical federated learning challenges (e.g., non-iid data)?

Key findings

• FCF converges to the same optimum as standard CF (differences approach 0% after sufficient SGD steps).
• On MovieLens, In-House, and simulated data, CF and FCF have nearly identical performance across metrics with average diff% < 0.5%.
• Bayesian correlated t-tests show high probability (≈0.999) that CF and FCF are statistically equivalent in precision within ±0.5%.
• Adam-based adaptive learning stabilizes convergence across a wide range of implicit confidence α values.
• FCF preserves privacy by keeping user data on-device and aggregating gradient updates without identity information.