FLRF: Federated recommendation optimization for long-tail data distribution

Abstract

Recommendation systems play a crucial role in real-world applications. Federated learning allows training recommendation systems without revealing users’ private data, thereby protecting user privacy. As a result, federated recommendation systems have gained increasing attention in recent years. However, The long-tail distribution problem of federated recommendation systems has not received enough attention. A small number of popular items receive most of the users’ attention, while a significantly larger number of less popular items receive feedback from only a small portion of users. Existing federated recommendation systems usually train on datasets with a long-tail distribution, which can easily lead to over fitting on a small number of popular items, reducing the diversity and novelty of recommendations and causing popularity bias. This paper proposes FLRF, a Federated Long-tail Recommendation Framework, which consists a long-tail recommendation model based on disentangled learning and a long-tail-aware aggregation method based on the attention mechanism. The long-tail recommendation model utilizes the idea of disentangled representation learning to explicitly disentangle the attractiveness of items into fame and niche. The long-tail-aware model aggregation, performs global attention aggregation on the model parameters of the fame part and self-attention aggregation on the model parameters of the niche part. We conduct comparative experiments on the three real-world datasets against the baseline methods in terms of accuracy and novelty. The experimental results show that the proposed framework can improve the diversity and novelty of recommendations without significantly impacting recommendation accuracy.