Multi-Extraction Techniques of Content-Based Music Recommender System

This study addresses the challenge of music rec-ommendation in the digital age, where the diversity of user preferences necessitates an effective filtering method. Traditional approaches like collaborative filtering struggle with new users due to a lack of item history, and content-based methods often rely solely on textual attributes like lyrics and genres, neglecting the rich information within audio files. Previous works have been limited by a narrow set of music feature extraction techniques. In contrast, this research leverages a multi-extraction approach to enhance recommendation accuracy by utilizing six distinct audio feature extraction methods-MFCC, zero crossing rate, spectral centroid, spectral bandwidth, spectral roll-off, and chroma STFT. Cosine similarity is used to determine the similarity between audio files. Utilizing the FMA dataset and evaluating the system with the NDCG method, the study achieves a notable score of 0.957, demonstrating the high relevance of the music recommendations provided. This multi-extraction technique approach represents a significant advancement in audio-based music recommendation systems.