Exploring different item embeddings for music recommender systems

Abstract

ABSTRACT: The proliferation of streaming services has led to an unprecedented increase in the availability of music, movies, and books, which has brought the need to develop sophisticated recommender systems to assist users in navigating this vast amount of items. Recommender systems play a crucial role in suggesting items to users based on their short-term interactions, in particular When long-term user data is unavailable, or user sessions are anonymous. This research adresses a gap in the literature by comparing the performance of various embedding methods for item representation in recommender systems. Embedding techniques transform items into vectors in a low-dimensional space, capturing semantic and structural relationships that are crucial for accurate recommendations. Thus, the main goals of this research are to generate independente sets of embeddings using popular techniques, optimize recommendation algorithms, and evaluate the embedding methods on recommendation performance. In this work, we intend to evaluate the effectiveness of different embedding methods, including Word2Vec, DeepWalk, Node2Vec, and LINE, in enhancing the performance of session-based recommender systems of increasing complexity, from simple arithmetic mean models to advanced neural network architectures such as Long Short-Term Memory (LSTM) networks and Transformers. We evaluate our work in terms of Precision, Recall, F1-score, Mean Average Precision (MAP), and Normalized Discounted Cumulative Gain (NDCG); by running an empirical evaluation in two different music datasets: Music4All and Xiami. Additionally, we examine the coverage of recommendations to ensure a balanced representation of both popular and niche items, addressing the long-tail effect in which a small number of items dominate user interactions. The expected outcomes include identifying embedding methods that consistently improve the performance of recommendation models across different complexities and datasets in the music domain. By providing a detailed comparative analysis, this research offers valuable insights into the selection and application of embedding techniques for developing effective recommender systems in this specific domain.