Cross-Domain Cold-Start Recommenders for Product Discovery

Recommendations are a core component of modern personalized e-commerce experiences, yet designing relevant recommendations for new users with little or no interaction history remains challenging. This cold-start setting becomes even more difficult in rapidly evolving product categories, where new items are introduced frequently and interaction data is sparse. To address this problem, we propose a cross-domain recommendation framework that leverages user behavior from related product domains together with item metadata to improve recommendations in a target domain with limited history. Our approach uses hybrid factorization methods that represent users and items through combinations of content and interaction-derived latent factors, enabling effective recommendation in both user and item cold-start settings. We show that the proposed method outperforms collaborative-only and content-based baselines under sparse-data conditions, while remaining competitive when interaction data is abundant. In addition, the learned feature embeddings capture meaningful semantic relationships that are useful for related recommendation tasks.