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.
Hanuma Ramesh Chadalavada (Fri,) studied this question.