One of the main challenges in recommender systems is the cold-start problem, in which recommendation systems struggle to recommend new or rarely visited items. The traditional methods usually comprise centralized data merging or collaborative filtering techniques, which are not easily applicable in the decentralized settings. The current federated recommendation techniques like FedMF and FedGN have limited support for cold-start personalization, particularly in situations where the metadata of the items is sparse or non-existent. To overcome these drawbacks, we propose a new federated learning-based model, CRAFT (Cold-start Recommender with Attention and Federated Training), that improves cold-start recommendations without compromising the privacy of the user. CRAFT proposes an attention mechanism to highlight salient user-item interaction patterns to enhance the inference of user preferences. Every client then trains a personalized model locally, where the updates are collectively aggregated through Federated Averaging (FedAvg) so that the collective intelligence is obtained without losing the sensitive information. CRAFT provides very personalized suggestions by adding time-varying dynamics and rich interaction histories. CRAFT can also be scaled to be deployed across distributed environments with the use of NVFlare platform. As indicated by experimental results on three real world datasets, including MovieLens 1M, Amazon Movies & TV and CiteULike, CRAFT is able to achieve nDCG 20 in cold-start scenarios up to 16.8 better than state of the art baselines, with strong privacy guarantees.
Sivakumar et al. (Tue,) studied this question.