Cultural differences in music features across Taiwanese, Japanese and American markets

Cultural differences in music preference

Typically, most of the research in cultural differences in the psychological literature has come from top-down, theoretical approaches. These have been instrumental in shaping the field, by increasing awareness of systematic ways by which people from cultures are different. One of the most instrumental differences is in the independence and interdependence of the self (Markus & Kitayama, 1991, 2010). Westerners generally tend to be independent, in that they prioritise the autonomy and uniqueness of internal (self) attributes. In contrast, East Asians generally tend to be interdependent, where their concept of self is intricately linked to close social relationships. This has been shown to have implications on music preferences through differences in desirability of emotions. For example, Westerners generally tend to view happiness as a positive and internal hedonic experience to be maximised where possible (Joshanloo & Weijers, 2014). As such, Western music preference tends towards high-arousal music, possibly in the search of strong ‘happiness’ experiences. East Asians, however, view happiness as a positive feeling associated with harmonious social relationships, in contrast to the hedonic, high-arousal definition in Western contexts. Consistently, East Asian preferences for music do not have this high-arousal component, and is more calm, subdued and relaxing (Tsai, 2007; Uchida & Kitayama, 2009; Park et al., 2019).

However, such theoretically based analyses may overrepresent Western cultures in research and literature. Consequently, cultural differences within similar, non-Westernised spheres are not well understood, due to the lack of pre-existing theory. For example, Chinese and Japanese cultures are often grouped together in cross-cultural research as a representation of East Asian collectivism, that functions as a comparative antithesis to Western findings (e.g., Heine & Hamamura, 2007). Yet, research has also uncovered differences between China and Japan that cannot be explained by these theories (Muthukrishna et al., 2020). As such, cultural differences within East Asia are not well understood in the psychological literature, and few theories exist to offer predictions on differences in music preference between these cultures.

Our solution was to examine music as cultural products from the bottom-up. Doing so would reduce the effect of experimenter bias in guiding theory formation and interpretation, when examining a wide database of music features. Cultural products are behavioural manifestations of culture that embody the shared values and collective aesthetics of a society (Morling & Lamoreaux, 2008; Lamoreaux & Morling, 2012; Smith et al., 2013). This implies that music consumption behaviour underscores culturally based attitudes, cognitions and emotions that afford preferences for certain congruent types of music. For example, in a cross-cultural comparison between Brazil and Japan, De Almeida & Uchida (2018) found that Brazilian song lyrics contained higher frequencies of positive emotion words and lower frequencies of neutral words than Japanese lyrics. This was consistent and reflective of their respective cultural emphases on emotion expressions (see Triandis et al., 1984; Uchida & Kitayama, 2009), and showed that comparing music ‘products’ elucidated differences between the collective shared values of different cultures. Past research on cultural products have relied on both popularity lists (charts; e.g., Askin & Mauskapf, 2017), and on artifacts produced by a culture (e.g., Tweets: Golder & Macy, 2011; newspaper articles: Bardi, Calogero & Mullen, 2008). We utilise both methods, and propose that examining cultural differences in ‘music’ products on a large-scale may provide potential insight into the sociocultural circumstances that give rise to these differences.

The present research

We adopted a data-driven, bottom-up approach to explore music preferences between cultures/industries through musical features for this study through machine learning. i.e., we first train a multiclass model to classify songs as belonging to (originating from) Chinese (Taiwanese), Japanese, or Western (American/English) markets. This is to establish the presence and magnitude of discernible cultural differences in music features. Next, we decompose the model by training three binary machine learning classifiers to classify songs as belonging to one culture or another. By applying model interpretation techniques on these models (such as Partial Dependence Plots (PDPs)), we aim to discover the specific difference in preferred musical features between Chinese (Taiwanese)-Japanese markets, Chinese (Taiwanese)-Western (American/English) markets and Japanese–Western (American/English) markets. We aimed to include as many songs as possible that were produced from these respective culture-based music industries to observe systematic trends and differences from as wide a range of musical styles and genres within these industries as possible. Finally, we examine the generalizability of these interpretations by conducting a follow-up study on Top-50 songs from Taiwan, Japan, and the US. If the identified features of difference present in songs produced by a cultural market were indeed representative of cultural differences in music preferences, we expect that these features should also show consistent differences for their respective Top-50 (popular) songs.

Overview

To explore differences between cultures, we used machine learning to classify a database of Chinese, Japanese, and English songs into their respective cultural (linguistic) markets. As this was a multiclass classification problem, we conducted the analysis twice using gradient boosted decision trees (GBDTs) and artificial neural networks (multi-layer perceptron, MLP) that are inherently capable of multiclass classification. This was also to examine the consistency in results between two differing methods of analysis, and strengthen the reliability of the analyses. To infer the features that accounted for cultural differences, we use model interpretation techniques, namely relative feature importance (RFI; Friedman, 2001), permutational feature importance (PFI; Fisher, Rudin & Dominici, 2018), and partial dependence plots (PDPs, Friedman, 2001), to examine and visualise the relationships between the feature and its influence on the probability of classification.

Data mining

We accessed the Spotify Application Programme Interface (API) through the ‘spotifyr’ wrapper (Thompson, Parry & Wolff, 2019) in R, to obtain song-level music feature information from Chinese, Japanese and English artists from the Spotify database. This was through a pseudo-snowball sampling method: we relied on Spotify’s recommendation systems (the ‘get_related_artists’ function) to recommend artists related to those in the official Spotify Top-50 chart playlists for Taiwan, Japan, and the US respectively and created a list of artists per country. We then used the same method to obtain another list of recommended artists to these respective ‘lists’, for up to six iterations, in order to obtain comparable sample sizes between these three markets. We also excluded all non-Chinese, non-Japanese, and non-English (language) artists from the respective list. This was through an examination of the associated genres for each artist, which often contained hints to their cultural origins (e.g., J-pop, J-rock, Mandopop). Artists that did not have listed genres were checked manually by the researchers. This resulted in a final N(artists) = 10,259 (Japanese = 2,587; English = 2,466; Chinese = 5,206). All song-level feature information for all artists were then obtained from the Spotify database. Duplicates (such as the same song being rereleased in compilation albums) were removed, for a total of N(songs) = 1,810,210 (Japanese = 646,440; Chinese = 360,101; English = 803,669). To ensure class balances, we randomly downsampled the Japanese and English samples to match the Chinese sample, resulting in a final N(songs) = 1,080,303.

Data handling and analysis

Except for ‘key’ and ‘time signature’, all Spotify features were inputted as features in the classification models. These were: ‘danceability’, ‘energy’, ‘loudness’, ‘speechiness’, ‘acousticness’, ‘instrumentalness’, ‘liveness’, ‘valence’, ‘tempo’, ‘duration’ (ms), and ‘mode’. A list of definitions for these features is available in Table 1. These were to classify songs according to their cultural membership (Chinese, English, or Japanese), as the outcome variable. The data was split into a training and testing set along a 3:1 ratio. Parameters for the GBDT model and weights for the MLP model were tuned through 5-fold cross validation on the training set. We also examined RFI scores for each model. For GBDT, this was a measure of the proportion that a feature was selected for stratification in each iterative tree, and for MLP, this was based on PFI, which measures the resultant error of a model when each feature is iteratively shuffled—the greater the error, the larger the influence a feature exerts on the outcome variable (Fisher, Rudin & Dominici, 2018; Molnar, 2019). We then simplified the classification problem by splitting it into 3 separate binary classifications: Japanese–Chinese, Japanese–English and Chinese–English. GBDTs and MLP models were conducted for these three comparisons, and in addition to RFI measures, we visualised the effect of each variable using PDPs. These show the averaged marginal effect of a feature on the outcome variable in a machine learning model, and is useful to glean an understanding of the nature of the relationship between these variables. PDPs were conducted through the ‘pdp’ package (Greenwell, 2017), and PFIs were conducted through the ‘iml’ package (Molnar, Bischl & Casalicchio, 2019). Machine learning was conducted through the ‘gbm’ package (Greenwell et al., 2019) for GBDTs, and the ‘nnet’ package (Venables & Ripley, 2002) for MLPs, via the ‘caret’ wrapper (Kuhn, 2019) in R (R Core Team, 2019). All R scripts used for data mining and analysis are available in our OSF repository (https://osf.io/d3cky).

Descriptives

Table 2 reports the descriptive medians, lower/upper quantiles, and missing data for each feature per culture. The full list of artists, genres, and songs are available in our OSF repository (https://osf.io/d3cky). Additionally, we note that while our database of songs spans as early as the 1950s, most of the songs in our database were from the mid-2000s to 2020 (see Fig. 1).

Multiclass classification (Chinese–Japanese–English)

For the GBDT model, the parameter tuning resulted in N(trees) = 150, interaction depth = 3, alongside default parameters of shrinkage = 0.1, and number of minimum observations per node = 10. The GBDT achieved a classification accuracy of 0.682, 95% CI [0.680–0.683], significantly above the no information rate (NIR) of 0.333, p < 0.0001. Aside from the input and output layers, we used a MLP model consisting of one hidden layer with five nodes. The MLP model achieved a slightly lower accuracy score of 0.660, 95% CI [0.659–0.662], but was still significantly above the NIR of 0.333, p < 0.0001. RFIs for both models are reported in Table 3.

Binary classifications

We first unpack the Chinese–Japanese model: For the GBDT model, the parameter tuning resulted in N(trees) = 150, interaction depth = 3, and no changes were made to the other default parameters (as above). The GBDT achieved a classification accuracy of 0.784, 95% CI [0.783–0.786], AUC = 0.865, significantly above the no information rate (NIR) of 0.500, p < 0.0001. The MLP model achieved a comparable accuracy score of 0.766, 95% CI [0.764–0.768], AUC = 0.844, significantly above the NIR of 0.500, p < 0.0001. Next, the Chinese–English model: For the GBDT model, the parameter tuning resulted in N(trees) = 150, interaction depth = 3, and no changes were made to the other default parameters. The GBDT achieved a classification accuracy of 0.807, 95% CI [0.805–0.809], AUC = 0.885, significantly above the no information rate (NIR) of 0.500, p < 0.0001. The MLP model achieved a comparable accuracy score of 0.803, 95% CI [0.801–0.805], AUC = 0.880, significantly above the NIR of 0.500, p < 0.0001. Finally, the Japanese–English model: For the GBDT model, the parameter tuning resulted in N(trees) = 150, interaction depth = 3, and no changes were made to the other default parameters. The GBDT achieved a classification accuracy of 0.713, 95% CI [0.711–0.715], AUC = 0.797, significantly above the no information rate (NIR) of 0.500, p < 0.0001. The MLP model achieved a comparable accuracy score of 0.709, 95% CI [0.707–0.711], AUC = 0.791, significantly above the NIR of 0.500, p < 0.0001. All RFIs and PFIs are reported in Table 4. Additionally, the two most important features are visualised by PDPs in Fig. 2. A visual inspection of the PDPs suggests that English music is higher than both Japanese and Chinese music in speechiness, Chinese music is higher than both Japanese and English music in acousticness, and Japanese music is higher than English and Chinese music in energy. In comparing Japanese and Chinese music, we note that acousticness and energy were also present, but were identified only in the GBDT model. In contrast, the MLP model identified loudness and instrumentalness as higher in Japanese music than Chinese music.

Overall, this suggests that, unlike English–Japanese or English–Chinese comparisons which were markedly different on a few main features, the differences between Chinese and Japanese music were spread widely across the various features. Consequently, despite relying on different ‘important variables’, both the MLP and GBDT managed to achieve a comparably high classification accuracy, with the GBDT outperforming the MLP for all classification tasks (results from DeLong’s tests are available on our OSF repository).

Additional analyses

We also visualised the changes in features over time for speechiness, acousticness, energy, instrumentalness, and loudness, from 2000 to 2020 (Fig. 3). Feature information for songs before 2000 were excluded due to the markedly smaller sample. Other than instrumentalness, which showed a notable decrease over time in Japanese songs, the remaining four features showed stability over time. This suggests that the differences in preference highlighted by the RFIs, PFIs and PDPs could indicate long term cultural preferences for music.

Follow-up study

We obtained a second round of data (approximately one year later) from Top-50 lists for Japan, Taiwan and the USA. Focusing on the identified features of speechiness, acousticness, energy, instrumentalness, and loudness from the previous study, Kruskal-Wallis tests revealed a significant effect of energy on speechiness (χ²(2) = 30.5, p < 0.001), acousticness (χ²(2) = 24.5, p < 0.001), energy (χ²(2) = 21.0, p < 0.001), and loudness (χ²(2) = 33.5, p < 0.001), but no significant effect was observed for instrumentalness (χ²(2) = 1.4, p = 0.49). For speechiness, post-hoc Dwass-Steel-Critchlow-Fligner pairwise comparisons revealed that USA was significantly higher than Taiwan (W = 7.34, p < 0.001) and Japan (W = 5.91, p < 0.001), but no significant difference was observed between Taiwan and Japan (W = -1.48, p = 0.55). For acousticness, Taiwan was significantly higher than Japan (W = 5.97, p < 0.001) and the USA (W = 6.11, p < 0.001), but no significant difference was observed between the USA and Japan (W = 0.53, p = 0.93). For energy, Japan was significantly higher than Taiwan (W = 6.01, p < 0.001), and the USA (W = 4.35, p = 0.006), but no significant difference was observed between Taiwan and the USA (W = 2.88, p = 0.103). For loudness, Japan was significantly higher than Taiwan (W = 7.44, p < 0.001) and the USA (W = 6.36, p < 0.001), but no significant difference was observed between Taiwan and the USA (W = 2.20, p = 0.27). Finally, for instrumentalness, no significant difference was observed between Japan and Taiwan (W = 0.61, p = 0.90), Japan and the USA (W = 1.67, p = 0.48), or Taiwan and the USA (W = 1.01, p = 0.75).

In short, with the exception of instrumentalness, Top-50 playlists obtained one year later nevertheless demonstrate strong consistency with the earlier results. American Top-50 songs are higher than both Japanese and Taiwanese Top-50 songs in speechiness, Taiwanese Top-50 songs are higher than both Japanese and American Top-50 songs in acousticness, and Japanese Top-50 songs are higher than American and Taiwanese music in energy and loudness. However, instrumentalness, that was originally identified as a variable of importance for the MLP model, did not consistently differ between cultures. Indeed, Fig. 3 shows that instrumentalness in Chinese music is inconsistent, with strong fluctuations depending on year. More research is needed to determine if instrumentalness is indeed a preferred feature in Taiwanese markets or merely a passing trend.