Cluster Analysis

In order to find the clustering relationship among different sound features, we plot three bubble-scatter plot to see if there is a clustering.

Bubble plots on selected audio features

In the cluster analysis, we aim to find out are there sub-categories living in the given categories, using “pop” as an illustration. In order to find the clustering within “pop”, we performed clustering analysis using k means clustering, hierarchical clustering, and DBSCAN. By the k-means distance plot and the silhouette coefficients, we observed a negative relationship between clustering quality and the number of cluster.

Based on the silhouette scores, hierarchical clustering with 2 final clusters produces clusters with best quality. However, overall, k-means clustering has a higher average silhouette score and a better fit. The highest silhouette is around 0.33, indicating that there may not be any subcategory.

• K-Means

From the following cluster plot, we can see that when k = 2, there is a clear distinction among two clusters. When the number of cluster increases, the shape and boundary of each cluster get blurred, showing suboptimal quality.

• Hierarchical Clustering

The hierarchical clustering plots show the similar result to k-means. When n = 2, the two clusters have clear boundaries. When n = 3, there are mixed up among clusters. When n = 5 and 20, all the points are mixed up and there is no clear distinction among clusters.

• DBSCAN

Using dbscan, we only get one cluster.

Result

Overall, the data(“pop”) has weak subcategories. From the 3-D clustering plots, we observe two potential subcategories formed by songs with higher energy, danceability, and acousticness, and their counterparties.