Capturing User Demands from Sparse Interactions
Overview
We are committed to advancing AI-driven recommender systems that make online services more accessible, personalized, and inclusive for all users. While personalization has long been a cornerstone of digital streaming, a persistent challenge remains: how to recommend relevant content to new or inactive users who have little or no viewing history. These “cold-start” users represent a large portion of real-world audiences – people who are new to a platform or engage intermittently, often receive less satisfying recommendations, limiting their ability to discover content that truly reflects their interests.
In a PhD internship project with research student Xurong (Jason) Liang and Amazon, the cross-institutional team leverages large language models (LLMs) to simulate meaningful interaction histories for cold-start users by interpreting simple demographic attributes. The model generates a set of imaginary yet plausible video interactions, which are then incorporated into sequential recommender systems to enrich user personalization profiles. This innovation brings real-world value to large-scale streaming platforms where equitable personalization is essential. By enriching the experience for users who lack historical data, it ensures that new customers, regardless of their background or data availability, are not disadvantaged in accessing high-quality, relevant content. Experiments on Amazon’s internal streaming service data reveal a significant boost (13.7% using NDCG@20) in recommendation quality for cold-start users. This solution also reduces biases in recommendation pipelines that often overfit to frequent users, thereby supporting a more balanced ecosystem of content discovery. By bridging generative AI and recommender systems, this research outcome showcases how enterprise-scale machine learning can deliver both measurable business impact and tangible social benefit by helping users feel recognized and understood, even from their very first interaction.
