1. Constructing a Corpus

In order to better understand the potential differences between Western and Korean pop music, I aimed to create a fair depiction of what is considered to be the most “popular” in both genres from a data analytics perspective. Many musical platforms such as Billboard and Apple Music provide musical rankings in popularity, but I found Spotify to be the most accurate in depicting popularity in terms of geographic and personalized streaming preferences. Thus, popularity was operationalized by stream count from Spotify’s API data, including track and artist popularity provided by the End-of-Year Wrapped from Spotify Newsroom. 

The study conducts a comparative corpus study on two datasets: the Top Hits in the USA (2020-2024) and the Top K-Pop Hits (2020-2024) on Spotify. Initially, the scope was restricted to the Top 50 Hits in the USA and the Top 50 K-Pop Hits in 2022; however, this was expanded as the initial corpus proved insufficient for discerning significant insights. The current corpus consists of 165 songs in each corpus, made up of the top 10 to 50 songs from each year, and the most popular songs in both categories as of April 2024 (See Appendix A). While recognizing the limitations of defining Western music as “American,” I decided to not utilize the Top Global Hits as the Western corpus since there were K-Pop songs in those playlists that could potentially confound the results. 

2. Computational Methods 

To address my research question, I programmed Python scripts to extract the track ID, track key, key confidence, and mode confidence for the Top 50 songs in the USA and K-Pop (2022) corpus. Based on the track IDs, I plotted key plots using Albrecht and Shanahan’s key-finding algorithm with humdrumR and compmus R packages (2013). The original graphs were created using Krumhansl-Schmuckler Key-Finding Weightings to serve as the basis for comparison. Then, I generated a K-Pop Major and Minor Key Profile that was trained on the K-Pop corpus, and a Western Major and Minor Key Profile trained on the Western corpus. The three key-finding weightings were used to visualize selected tracks and compare how successful each key profile was at predicting their keys. Then, I created a K-Pop classifier, employing a neural net trained on the features generally associated with K-Pop, and then tested the accuracy of that model using R scripts. The audio features I chose to focus on were acousticness, valence, loudness, liveness, danceability, speechiness, and mode confidence. 

To better understand the model fit for each audio feature of interest, I used the Akaike Information Criterion (AIC) measure to gauge which specific group of audio features is most relevant to predicting K-Pop. The AIC is defined as a summary statistic used in comparing the relative goodness of fit of two or more models for a given set of data, while taking into account the number of parameters in each model. The model with the lowest AIC is considered the best among all models specified (APA, 2018).

Based on the findings of AIC, I conducted a regression tree analysis to assess how the model distinguishes what is considered “K-Pop”. I compared the original neural network model against a logistic regression (glm), K-nearest neighbor (knn), learning vector quantization (LVQ), gradient boosted machine (GBM), support vector machine (SVM), and random forest (rf). The model then produced a summary of its resamples for each dataset. In addition to the regression tree, I also conducted a logarithmic odds ratio analysis to understand each audio feature variable’s predictive value in classifying a song as K-Pop. The increase in the corresponding predictor variable is associated with an increase in the log odds of the song being classified as K-pop, while a negative coefficient indicates the opposite.