Department of Computer Engineering

The rapid growth in the diversity and volume of digital content has intensified the challenge of efficiently delivering relevant items to users. Recommendation systems have emerged as an effective means of addressing this challenge, enhancing user satisfaction while improving engagement and retention for platforms. In recent years, the integration of large language models (LLMs) into recommendation systems has gained increasing attention due to their ability to perform tasks such as knowledge injection, pattern recognition, and user profiling. This thesis introduces Real World Recommender, a hypernetwork-based architecture that leverages human-readable user profiles generated by LLMs to construct personalized recommendation networks for individual users. Within this framework, the hypernetwork produces the weights of a secondary neural network responsible for predicting user preferences based on the generated profiles. By tailoring a dedicated neural network to each user, the proposed approach offers a novel utilization of LLM-based user representations in personalized recommendation systems. Experimental evaluations conducted on widely used benchmark datasets demonstrate that the method consistently outperforms established baselines, highlighting its effectiveness in generating accurate and well-ranked recommendations.