When Federated Recommendation Meets Cold-Start Problem: Separating Item Attributes and User Interactions

TL;DR

The paper tackles cold-start recommendations in federated learning by introducing IFedRec, which maintains two item representations: one on clients derived from user interactions and another on the server derived from item attributes via a meta attribute network. An item representation alignment mechanism bridges these representations, enabling effective cold-item recommendations without exposing private interactions or raw attributes. The framework uses a two-phase learning process on warm items and inference on cold items, with optional Local Differential Privacy to enhance privacy. Empirical results on four cold-start benchmarks show state-of-the-art performance and robustness to limited client participation and noise, suggesting practical applicability for privacy-preserving federated recommendation systems.

Abstract

Federated recommendation system usually trains a global model on the server without direct access to users' private data on their own devices. However, this separation of the recommendation model and users' private data poses a challenge in providing quality service, particularly when it comes to new items, namely cold-start recommendations in federated settings. This paper introduces a novel method called Item-aligned Federated Aggregation (IFedRec) to address this challenge. It is the first research work in federated recommendation to specifically study the cold-start scenario. The proposed method learns two sets of item representations by leveraging item attributes and interaction records simultaneously. Additionally, an item representation alignment mechanism is designed to align two item representations and learn the meta attribute network at the server within a federated learning framework. Experiments on four benchmark datasets demonstrate IFedRec's superior performance for cold-start scenarios. Furthermore, we also verify IFedRec owns good robustness when the system faces limited client participation and noise injection, which brings promising practical application potential in privacy-protection enhanced federated recommendation systems. The implementation code is available

Figures (5)

• Figure 1: Three cold-start recommendation systems comparison. The centralized method (a) saves raw item attributes on the server but exposes private user interaction records. Traditional FedRecSys (b) secures the interaction records but exposes the item attributes to the clients. Our IFedRec (c) can protect these two types of security-sensitive information.
• Figure 2: The framework of IFedRec. During the learning phase, the client uploads the item embedding to the server for global aggregation, and other recommendation modules are preserved locally to capture user personalization. On the server side, we elaborate a meta attribute network to learn item attribute representation based on raw item attributes. Besides, an item representation alignment mechanism is developed to align two item representations, i.e.,$\mathcal{L}_{global}$ and $\mathcal{R}$. During the inference phase, the server first learns the cold item attribute representation, and then each client can make personalized recommendations by integrating it with locally preserved recommendation modules.
• Figure 3: Impact of the regularization coefficient. The horizontal axis is the value of the regularization coefficient $\lambda$, and the vertical axis is the Recall metric.
• Figure 4: Impact of the meta attribute network training epoch. The horizontal axis is the value of meta attribute network training epoch $E_1$, and the vertical axis is the Recall@20.
• Figure 5: Convergence analysis about the client amount participated in each communication round. The horizontal axis is the client sampling ratio, and the left vertical axis is the number of communication rounds, the right vertical axis is model performance on three metrics.