In extreme cold-start scenarios, where users and items have little or no interaction history, traditional recommender systems struggle to generate diverse and meaningful suggestions. This work introduces a unified framework designed to address this challenge using two complementary techniques: (i) a graph augmentation module which densifies the user-item graph using feature-based link prediction, and (ii) a re-ranking algorithm which increases recommendation diversity via centroid-aware sampling across distant user clusters. The resulting architecture also combines graph neural networks for representation learning. We empirically evaluate the framework using a real-world industrial dataset from the telecommunications sector, which is characterized by severe data sparsity and item imbalance. A comprehensive exploration of the possible hyperparameter configurations is conducted, allowing for the assessment of diversity-coverage trade-off and enabling the selection of a configuration that meets specific objectives. The results provide actionable insights for practitioners, demonstrating how to optimize the framework according to specific goals, whether prioritizing diversity, coverage, or computational efficiency. This work serves as a resource for deploying scalable high-quality recommendation systems in real-world industrial settings, addressing the unique challenges posed by extreme cold-start scenarios.
Sbandi et al. (Mon,) studied this question.