Multi-Selection for Recommendation Systems

2025-04-10

AI Summary

To address the inherent trade-off between local differential privacy (LDP) and recommendation utility, this paper proposes a multi-choice recommendation paradigm: the server generates multiple candidate items and deploys a lightweight local model to users’ devices, enabling clients to autonomously select the optimal item based on their private features. This approach is the first to jointly optimize privacy at both the server-side (global modeling) and client-side (personalized decision-making) under LDP constraints. Under ε ≈ 1, it achieves 97% of the utility of the optimal non-private baseline—outperforming conventional single-choice mechanisms by six percentage points. Technically, the framework integrates deep neural networks (trained on MovieLens-25M), LDP-compliant perturbation, on-device lightweight inference, and a novel multi-candidate generation and ranking mechanism, significantly improving the privacy–utility equilibrium in multi-choice recommendation scenarios.

Technology Category

• Machine Learning: Privacy
• Search and Optimization: Local Search
• Data Mining & Knowledge Management: Recommender Systems

Application Category

• User Modeling, Personalization and Recommendation: User privacy protection in personalized systems
• Security and Privacy: Privacy-enhancing technologies
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We present the construction of a multi-selection model to answer differentially private queries in the context of recommendation systems. The server sends back multiple recommendations and a ``local model'' to the user, which the user can run locally on its device to select the item that best fits its private features. We study a setup where the server uses a deep neural network (trained on the Movielens 25M dataset as the ground truth for movie recommendation. In the multi-selection paradigm, the average recommendation utility is approximately 97% of the optimal utility (as determined by the ground truth neural network) while maintaining a local differential privacy guarantee with $epsilon$ ranging around 1 with respect to feature vectors of neighboring users. This is in comparison to an average recommendation utility of 91% in the non-multi-selection regime under the same constraints.

Problem

• Develops a multi-selection model for private recommendation queries
• Ensures local differential privacy while maintaining high utility
• Compares multi-selection vs non-multi-selection recommendation performance

Innovation

• Uses deep neural network for recommendations
• Implements local differential privacy guarantees
• Multi-selection model enhances recommendation utility