FedUTR: Federated Recommendation with Augmented Universal Textual Representation for Sparse Interaction Scenarios

A research team led by Kang Fu has introduced FedUTR (Federated Recommendation with Augmented Universal Textual Representation), a new method designed to solve a core weakness in privacy-preserving AI recommendations. Traditional federated recommendation systems rely heavily on a user's past interaction history (like clicks or purchases) to create item embeddings. This fails when data is sparse—a common 'cold-start' problem for new users or items. FedUTR's innovation is to incorporate a universal textual representation of items, such as product descriptions or titles, as a stable knowledge base that doesn't depend on any single user's history.

The system uses a Collaborative Information Fusion Module (CIFM) to blend this generic textual knowledge with a user's personalized interaction data locally on their device. A Local Adaptation Module (LAM) then fine-tunes the model to preserve individual preferences. For extremely sparse scenarios, a variant called FedUTR-SAR adds a sparsity-aware component to better balance universal and personalized information. The result is a more robust model that performs well even with little historical data, all while maintaining the privacy benefits of federated learning where raw data never leaves the user's device.

Extensive experiments demonstrate a significant leap in performance. When tested on four real-world datasets, FedUTR outperformed current state-of-the-art federated recommendation baselines by up to 59%. The paper also provides a theoretical convergence analysis, guaranteeing the method's effectiveness. This work, published on arXiv, addresses a critical bottleneck for deploying practical, privacy-first recommendation systems at scale, moving them beyond reliance on dense user histories.