Building User Profiles with Limited Interaction Data

Cold-start users—those with little or no prior interaction history—pose a significant challenge for recommender systems. This thesis investigates methods to quickly construct useful user profiles based on a limited number of initial interactions, such as the first few purchases or clicks. Using publicly available datasets, we explore simple techniques like weighted aggregation of item features and nearest-neighbor approaches to model new users' preferences. We evaluate how these early profiles impact recommendation quality compared to models requiring extensive user history and investigate the minimal number of interactions required to construct reliable user profiles.