Enhancing Music Data Clustering: An Empirical Analysis for Music Clustering and Scaling Optimization
Riz Lala, Gauravi Patankar, Aditya Nanasaheb Patil, Himani Deshpaande · 2023
This paper presents a deep dive into music recommendation and proposes MCSO (Music Clustering and Scaling Optimizer), a novel method to cluster songs and evaluate these clustered recommendations. Past research has explored different user-based techniques to predict and recommend suitable music. In this paper, songs are clustered based predominantly on their audio features using the MCSO approach. In MCSO, combinations of over 6 different clustering algorithms and 4 scaling techniques were evaluated using metrics such as the Silhouette score, the Calinski-Harabasz(CH) score, and the Davies-Bouldin score. The scores of these three indices indicate that among clustering algorithms, BIRCH, when used along with the Minimum Maximum Scaling technique produces the optimum result. The BIRCH clustering algorithm exhibited superiority over the second-best clustering algorithm (Mini-Batch KMeans) by 15.913% in the Silhouette Score, 8.577% in the Calinski-Harabasz Index, and 29.986% in the Davies-Bouldin Index. Additionally, amongst BIRCH itself, MinMaxScaler demonstrated a performance advantage of 0.264% in the Silhouette Score, 3.447% in the Calinski-Harabasz Index, and 0.816% in Davies-Bouldin Index over MaxAbsScaler and other top-performing Scaling techniques.