Використання великих мовних моделей для генерації синтетичних даних у рекомендаційних системах

Abstract

One of the main challenges in developing recommendation algorithms is the sparsity of user–item interaction data, which limits the effectiveness of traditional methods such as collaborative filtering. Synthetic data generation is considered a promising approach for expanding training datasets and improving the robustness of recommendation models. The proposed approach involves using a large language model to generate synthetic interactions between users and items based on descriptions of user preferences and item characteristics. Such generated data can simulate realistic usage scenarios and supplement existing datasets. The study proposes to investigate methods for generating synthetic interactions, modeling different types of user behavior, and evaluating the impact of generated data on recommendation quality.