Enabling Cold Start Recommendation Systems using LLM-Generated Synthetic Data : Developing Stock Basket Recommendation System in Data-Sparse Investment Environments

Sammanfattning

This thesis explores a novel approach to stock basket recommendation in investment platforms facing cold-start challenges due to a lack of user interaction history. By leveraging Large Language Models (LLMs) to generate synthetic user interaction data, a prototype recommendation system was developed using real-world financial metadata. Two recommendation strategies were implemented: content-based filtering using financial basket attributes, and collaborative filtering with models Singular Value Decomposition (SVD, a matrix factorization technique), K-Nearest Neighbors (KNN, a similarity-based algorithm), and deep learning models, trained on synthetically generated user-basket interactions. Evaluation using Top-K metrics (Precision, Recall, F1-score) showed that collaborative filtering models enhanced with LLM-generated synthetic data outperformed content-based filtering in recommendation quality, with the deep learning-based approach achieving the strongest results reaching a precision of approximately 60%. The results demonstrate that LLM-generated synthetic data can effectively mitigate the cold-start problem, offering a practical solution for emerging financial platforms aiming to deliver personalized investment recommendations without historical user data.