Application of LLMs and Embeddings in Music Recommendation Systems

Abstract

Music recommendation systems are crucial in guiding listeners to the extensive selection of music accessible today. However, these recommendation algorithms frequently encounter challenges such as the cold start problem, including less popular tracks, and comprehending the semantic substance of music. This thesis investigates the incorporation of Large Language Models (LLMs) and embeddings to tackle these difficulties in the context of music recommendation systems. This project utilizes two extensive datasets from Kaggle to create a hybrid recommendation model that blends content-based and collaborative filtering approaches with proficient LLMs.

The main goal is to make music suggestions more accurate and customizable. To do this, two methods are being used, which include adding embeddings to classic collaborative filtering techniques and using LLMs to add semantic analysis to content-based recommendations. These methods enable the system to catch the subtle preferences of users and suggest music that better matches their unique interests and moods.

The methodology includes preprocessing the datasets to make a single aggregrated dataset, using K-means clustering and Principal Component Analysis (PCA) to improve the representation of features, and using user interaction matrices to make collaborative filtering work better. The evaluation of the recommendation system is done by conducting user satisfaction surveys. The content-collaborative and LLM-based models have shown considerable performance, with an average satisfaction rating of around 4.2 out of 5.

This thesis discusses the implications of these findings for key research questions concerning the effectiveness of integrating LLMs and embeddings in music recommendation systems. The study provides a theoretical foundation for future research and prospective enhancements in music recommendation.