Abstract
BACKGROUND: Preferences for music can be represented through music features. The widespread prevalence of music streaming has allowed for music feature information to be consolidated by service providers like Spotify. In this paper, we demonstrate that machine learning classification on cultural market membership (Taiwanese, Japanese, American) by music features reveals variations in popular music across these markets. METHODS: We present an exploratory analysis of 1.08 million songs centred on Taiwanese, Japanese and American markets. We use both multiclass classification models (Gradient Boosted Decision Trees (GBDT) and Multilayer Perceptron (MLP)), and binary classification models, and interpret their results using variable importance measures and Partial Dependence Plots. To ensure the reliability of our interpretations, we conducted a follow-up study comparing Top-50 playlists from Taiwan, Japan, and the US on identified variables of importance. RESULTS: The multiclass models achieved moderate classification accuracy (GBDT = 0.69, MLP = 0.66). Accuracy scores for binary classification models ranged between 0.71 to 0.81. Model interpretation revealed music features of greatest importance: Overall, popular music in Taiwan was characterised by high acousticness, American music was characterised by high speechiness, and Japanese music was characterised by high energy features. A follow-up study using Top-50 charts found similarly significant differences between cultures for these three features. CONCLUSION: We demonstrate that machine learning can reveal both the magnitude of differences in music preference across Taiwanese, Japanese, and American markets, and where these preferences are different. While this paper is limited to Spotify data, it underscores the potential contribution of machine learning in exploratory approaches to research on cultural differences.