Feature-based User-centric Music Recommendation

Abstract

This paper presents a novel approach to music recommendation that leverages intrinsic musical properties to address the limitations of traditional collaborative filtering methods, particularly the cold-start problem. Unlike existing systems that rely heavily on user behavior and historical data, our proposed method utilizes a neural network trained on fundamental attributes of music—such as tempo, key, mode, duration, and loudness—to generate recommendations. We detail the process of extracting these musical properties using tools like the librosa library and a custom function, followed by training a neural network to predict user preferences based solely on these features. The system’s performance was evaluated using the Million Song Dataset, achieving an accuracy of 76.75%. Despite its potential, the method faces challenges related to computational efficiency and the handling of diverse musical genres. Future work includes exploring hybrid models combining collaborative filtering with feature-based recommendations and testing alternative algorithms to enhance performance and applicability.