Optimization of LightGBM for Song Suggestion Based on Users’ Preferences

Abstract

Undoubtedly, music possesses the transformative ability to instantly influence an individual's mood. In the era of the incessant flow of substantial data, novel music compositions surface on an hourly basis. It is impossible to know for an individual whether he/she will like the song or not before listening. Moreover, an individual cannot keep up with this flow. However, with the help of Machine Learning (ML) techniques, this process can be eased. In this study, a novel dataset is presented, and song suggestion problem was treated as a binary classification problem. Unlike other datasets, the presented dataset is solely based on users' preferences, indicating the likeness of a song as specified by the user. The LightGBM algorithm, along with two other ML algorithms, Extra Tree and Random Forest, is selected for comparison. These algorithms were optimized using three swarm-based optimization algorithms: Grey Wolf, Whale, and Particle Swarm optimizers. Results indicated that the attributes of the new dataset effectively discriminated the likeness of songs. Furthermore, the LightGBM algorithm demonstrated superior performance compared to the other ML algorithms employed in this study.