Universality of preference behaviors in online music-listener bipartite networks:

A Big Data analysis

Xiao-Pu Hana,b,c, Fen Linc , Jonathan J. H. Zhuc, Tarik Hadzibeganovicd

a Alibaba Research Center for Complexity Sciences, Hangzhou Normal University, Hangzhou 311121, China
b Institute of Information Economy and Alibaba Business School, Hangzhou Normal University, Hangzhou 311121, China
c Department of Media and Communication, City University of Hong Kong, Hong Kong, China
d Institute of Psychology, Faculty of Natural Sciences, University of Graz, Graz 8010, Austria

Abstract
We investigate the formation of musical preferences of millions of users of the NetEase Cloud Music (NCM), one of the
arXiv:2212.13139v1 [cs.SI] 26 Dec 2022

largest online music platforms in China. We combine the methods from complex networks theory and information sciences
within the context of Big Data analysis to unveil statistical patterns and community structures underlying the formation and
evolution of musical preference behaviors. Our analyses address the decay patterns of music influence, users’ sensitivity to
music, age and gender differences, and their relationship to regional economic indicators. Employing community detection in
user-music bipartite networks, we identified eight major cultural communities in the population of NCM users. Female users
exhibited higher within-group variability in preference behavior than males, with a major transition occurring around the age
of 25. Moreveor, the musical tastes and the preference diversity measures of women were also more strongly associated with
economic factors. However, in spite of the highly variable popularity of music tracks and the identified cultural and demographic
differences, we observed that the evolution of musical preferences over time followed a power-law-like decaying function, and
that NCM listeners showed the highest sensitivity to music released in their adolescence, peaking at the age of 13. Our findings
suggest the existence of universal properties in the formation of musical tastes but also their culture-specific relationship to
demographic factors, with wide-ranging implications for community detection and recommendation system design in online
music platforms.
Keywords: Musical preferences, complex networks, community detection, cross-cultural universals, gender differences, Big
Data.

1. Introduction music [19] and its developmental and evolutionary origins

[3, 12, 20, 21, 22, 23, 24, 25]. Individual musical prefer-
As a typical universal cultural feature [1, 2, 3, 4], music ences [26] can be seen as a bridge connecting these two per-
plays an important and irreplaceable role in our everyday spectives, rightfully attracting much recent attention across a
life [5]. Music not only provides pleasure, but it also con- number of disciplines [1, 27, 28, 29, 30, 31, 32, 33, 34, 35].
structs and shapes our social fabric [10]. It has been used However, previous investigations of musical preferences
widely for emotional regulation [6], pain management [7], have largely gravitated around the laboratory studies of indi-
enhancement of cognitive abilities [8], or for support of co- vidual music-listener behaviors in the offline world, whereas
ordinated movement behaviors such as dance [9]. Music has our understanding of musical tastes as they emerge and de-
evidently played a substantial role in the evolution of various velop online, in the context of modern social networking
cultural hallmarks, including the preservation of knowledge, technologies and music streaming platforms, remains largely
development of rituals and religious behaviors, as well as in limited.
the emergence of group cohesion that has been vital for our In the present paper, we study the development of musical
survival [5]. preferences in one of the largest online music communities
Due to its multidimensional nature, music can be studied in China. We combine the tools and methods from complex
from an array of different perspectives [10]. For instance, networks theory, information sciences, and physics within
one line of research focuses on musical features such as the context of Big Data analysis to unveil statistical univer-
rhythmic patterns [11], their statistical properties [12] and sality patterns of musical preferences and to detect the under-
influences on music composition [13], and their effects on lying listenership communities that emerge in online social
individual music appreciation [10, 14, 15]. Studies of this networks. We further provide quantitative characterization
kind have spawned a series of fruitful investigations into the of a number of variables related to musical preferences in
complexity of music [16] and its multi-leveled relationship to the studied population of online music listeners, including
human emotions and behavior [10, 15, 17, 18]. Another line their tagging diversity, community diversity, individual and
of investigation emphasizes the socio-cultural significance of global musical preferences, and the users’ overall sensitiv-
ity to music. We then model the obtained empirical data to
Email addresses for correspondence: [email protected] (Xiao-Pu identify relevant features of musical taste development, such
Han), [email protected] (Tarik Hadzibeganovic) as the age of users’ peak sensitivity to popular music. We
December 27, 2022
finally perform a series of analyses addressing the relation- ages, such as only adolescents and younger adults or only
ships between music preference variables and demographic middle-aged adult music listeners (for a discussion on this
factors such as gender, age, and economic development. issue, see Ref. [28]).
By applying network models to large-scale user-behavior Secondly, the vast majority of previous studies on musical
data from online music platforms, we show that global mu- preferences considered only individual-level offline perspec-
sical preferences can be captured and analyzed more effec- tives, whereas the direct investigation of musical tastes as
tively than what is usually attained in laboratory or field they emerge and evolve in social contexts, especially in on-
settings by assessing the habits and tastes of individuals or line social networks, is still lacking. Indeed, even though
small groups of music listeners. In addition, we demonstrate the study of online preference behavior has generally at-
that such network modeling and associated quantitative anal- tracted much recent attention [43, 44, 45], our understanding
yses of large-scale datasets can reveal statistical patterns and of users’ musical preferences as they emerge and develop
universal features of music identity formation, which would in online music-based communities is still largely limited
otherwise remain undetected through classical surveys or [1, 30, 32, 33]. Nevertheless, a more systematic study of
laboratory experiments. social networks in online music-streaming platforms and a
Our results are interpreted in terms of contemporary psy- deeper understanding of individual preferences in such on-
chological and socio-economic theories of musical taste de- line communities can be mutually revealing and highly infor-
velopment, as well as through the lens of recent technolog- mative for the design of novel music recommendation sys-
ical frameworks and complex systems approaches to user tems [46] or for the measurement of popularity and influence
preference behavior. Our findings thus contribute to a more of a wide variety of online contents [47, 48, 49].
integrative and multidisciplinary understanding of the devel-
opment of preference behaviors in complex social systems, 1.2. Methodological advances in Big Data analysis
which is relevant for a wide range of applications includ- The rise of large-scale online communities and Big Data
ing user profiling and community detection in online mu- technology have enabled access to huge amounts of infor-
sic networks, measurement of online content popularity, and mation. In recent years, the related large-scale data analyses
the design of playlist generation and music recommendation have been addressing nearly all aspects of social network-
systems for online social media. ing, including user behavior, system design, performance
analysis, or privacy issues, revealing a number of previously
1.1. Previous research and its limitations unknown trends and global patterns that remained largely
Employing personality questionnaires and surveys on mu- undetected when employing classical data evaluation tools
sical tastes, a series of studies have previously examined and techniques [19, 50, 51, 52, 53, 54, 55, 56]. Within this
individual differences in musical preference behaviors and context, especially in the study of music content popularity
the relationships between people’s taste in music and socio- in online social media, the network-theoretic structural ap-
psychological factors [27, 29, 36, 37]. Since personality and proach has gained considerable traction [19]. Importantly, by
social behaviors tend to vary as a function of demographic combining Big Data and network theoretic analyses of pref-
factors, many recent studies have investigated age trends and erence behaviors in online music-based communities, we
the associated temporal stability in the development of musi- have the unique opportunity to study not only the large-scale
cal preferences [28, 38, 39, 40]. Moreover, demographic fac- picture of musical preferences formation, but also to under-
tors such as age, sex, or economic status, have recently been stand how musical tastes are shaped by the psychological
viewed as even more influential than psychological factors as and social forces specifically emanating from the use of net-
predictors of genre-based musical preferences in laboratory working technologies and the associated computer-mediated
offline settings [28, 31]. interactions occurring among millions of users.
Nevertheless, despite offering rich insights into the ori- Coupling large-scale dataset analysis with network sci-
gins of musical preferences [27, 29], many earlier investi- ence can additionally allow us to paint a more global picture
gations have suffered consistently from several conspicuous of how various levels of sophistication of the online tech-
gaps. First, previous studies have focused mostly on veri- nology usage [57], such as online skills, users’ tagging be-
fying the existence of links between personality traits and havior, or the diversity of online activities, contribute to the
musical preferences, in spite of repeated unsatisfactory re- evolution of musical preferences. Excitingly, such studies
sults with rather small effect sizes (for a discussion, see e.g will undoubtedly offer integrative insights into the combined
[29]). Consequentially, these studies have often neglected effects of psychological, social, technological, and economic
other prominent aspects of musical preferences formation, factors on musical taste formation and change.
such as the global patterns of music popularity [19, 41, 42],
their variability with demographic factors, and their univer- 1.3. Research aims and hypotheses
sal features that may be shared across times and cultures. Since previous studies have mostly investigated tastes of
Survey-based methodologies or laboratory experiments are Western music listeners (e.g. the US American users of Spo-
limited in approaching such questions, as for example, due to tify, an online music streaming service), our present study
enormous costs, they are typically restricted to small sample focuses on the Big Data analysis of musical preferences of
sizes that are often not sufficiently representative of whole NetEase Cloud Music [58], one of the largest music stream-
populations and may therefore conceal the underlying global ing platforms and music-based online social networks in
trends. Critically, these small-sample studies are additionally China. The first aim of this study was to uncover age trends
limited by considering a rather narrow range of participants’ and age-related global behavioral patterns in musical taste
2
formation. By investigating these global features of musical
preferences in China, and by comparing them to those ob-
served with Western users of music-based social media, we
further aimed at identifying general properties and laws of
preference behaviors that may be shared universally across
distinct cultures and epochs. Since online one-to-many inter-
actions typically promote higher levels of cultural isolation
and diversity [59], finding such cross-culturally shared be-
havioral patterns in online music-based communities would
be a strong indicator of universality in the formation of mu-
sical preferences.

Specifically, different from earlier small-sampled labora-

Sex Ratio
tory studies [60], and in line with recent large-scale analyses
of online music streaming services in the US [61], we ex-
pected that the development of a listener’s sensitivity to mu-
sic in China would reach its peak prior to the early adulthood,
in the early adolescence of an individual. Thus, the strongest
influence on musical tastes of adult Chinese listeners should
occur already in their early teens, representing the first ma- Male

jor transition in the development of their musical preference

behavior. We derive this hypothesis from interactionist the-
ories (see e.g. [62]), postulating that changes in an individ-
ual’s development of musical preferences should be coupled
to the associated changes in basic psychological needs and Figure 1: (Color online)(a) The number of FP tracks as a function of the
release year. (b) The distribution of users’ age by gender. The numbers of
emotions, and to the emerging social identity formation of male and female users by age are denoted by blue and pink bars, respec-
an individual (also see [27, 28]). tively. The upper inset shows the sex ratio of users (the number of male vs.
female users) as a function of age, and the lower inset depicts the number of
songs in users’ FPs vs. age, separately for female and male users.
Next, with respect to the temporal stability of online mu-
sic influence, we expected to observe its slow-falling decay
that would mimic patterns typically found in the time series 2. Methods
of popular topics or in the outbreaks of memes in online so-
cial media [63, 64]. Furthermore, similar to earlier findings 2.1. Dataset
obtained with participants from the US and UK (e.g. [28]), NetEase Cloud Music (http://music.163.com, hereafter
we predicted that while interest in some musical dimensions abbreviated as NCM) is one of the world’s largest music-
would decrease with age in Chinese music consumers, their based online social networks [58]. Since its release in April
preferences for other musical dimensions would increase in 2013, hundreds of millions of users have registered with this
adulthood. music streaming platform. Using NCM software on a per-
sonal computer or smartphone, users can search, listen to,
Beyond this investigation of age trends in musical pref- or download over 10 million songs from its online music li-
erences, we addressed several specific assumptions related brary. Once registered with NCM, users can also build and
to gender differences in musical taste formation, their vari- edit playlists with their favorite songs, share their feelings
ability with age, and with regional economic development and comments, or recommend music in a Twitter-like online
indicators. More specifically, we expected that during ado- section of NCM, or on other online social media. Here, we
lescence, relatively to their male counterparts, female lis- use the term “songs” in a general sense to denote a piece of
teners would be characterized by more heterogeneous music music with or without lyrics, and we further interchangeably
preference behaviors that would also be stronger associated use the terms “song” and “track” throughout this article.
with economic indicators. Thus, musical preferences and The NCM dataset that was analyzed in the present study
attitudes should reflect gender and economic divides [31], was collected from the API of NCM by web crawlers in
particularly prior to adulthood of Chinese music listeners. the period from September 7 to September 27, 2016 1 . We
However, in accordance with the social role theory [65], we selected to crawl the first 200,000 playlist entries from all
predicted that this gender gap in the adolescence should de- available 2 million playlists. In total, we obtained a sam-
crease with the onset of the early adulthood and with the ple representing 10% of all of the NCM’s playlists, in which
increasing economic development, representing the second 30,562,590 playlists contained a total of 4,261,266 different
major transition in the development of musical preferences. tracks.
In addition, employing network theoretic analysis and the There were two types of playlists: One is called the Fa-
associated community detection algorithm, we expected to vorite Playlist (FP), which is created automatically when a
identify distinct communities of NCM users that would al-
low us to further investigate culture-specific musical prefer- 1 The dataset is freely available from the authors upon reasonable re-

ence behaviors at the community and group levels. quest.

3
user marks a track as a favorite for the first time. This list
contains all the marked tracks of the user. Each user is al-
lowed to create only one FP. The other type is the General
Playlist, which can be freely created and modified by users.
Each user can create no more than 1000 General Playlists.
Our analysis mainly focuses on FPs, because they contain
the music that a user listens to most frequently, thus reflect-
ing the user’s typical musical preferences. In total, we col-
lected 4,499,164 FPs, in which 3,028,351 FPs corresponded
to still active users, containing a total of 2,247,960 different
songs. The remaining tracks come from General Playlists,
amounting to 26,063,426 in total.
Each playlist contains detailed information about each
track included in the playlist, such as a song’s ID, title, its
corresponding album, album release date, URL information,
etc. As shown in Figure 1(a), users typically prefer new
Figure 2: (Color online) Schematic representation of the user-music bipar-
songs over old ones. Most favorite songs of NCM users were tite network and the calculation of the total attention A given to each song.
released in the last decade. The thickness of each gray line is proportional to the ratio of the user’s at-
Each playlist also contains the information about its user, tention focused on a given song.
such as gender, age, city, and province. However, we were
not able to access any information that could reveal user’s
users who have exceeded 25 years of age. The length of
personal identity. In our analyses, we thus only used the in-
the FPs – the number of favorite songs a user saves – also
formation that can be accessed publicly from NCM via web
changes with age. In the group of users below the age of
crawlers, and our data collection did not involve any interac-
25, the playlist length is similar among male and female lis-
tion with or manipulation of any research subjects.
teners, whereas among the users older than 25, males have
Figure 1(b) shows the age and gender distributions of
significantly longer lists of songs than their female counter-
NCM users that were collected for the present study. We
parts. In other words, NCM seems to attract more younger
see that most users were in their 20s. However, we note here
female users and more middle-aged male users.
that age, as well as other information, was self-reported by
Users can assign different tags to each General Playlist
users at the time of their account registration, and may thus
they created, with no more than three tags selected from
not always be completely accurate. Considering the reliabil-
a given tag-set of NCM. The tag-set includes five classes:
ity of age information, in the following analyses related to
Language, genre, scenario, emotion, and theme. Each class
age and gender, we only used the data of users aged between
further includes several tags, as shown in Table 1. Favorite
12 and 40, excluding all users who set their birthdate at Jan.
Playlists cannot be assigned any tags.
1, 1990 (which is the default setting of NCM). For this range
of ages (12 to 40), in each of the 29 age groups in our sam-
ple, we had at least 846 users. In 18 out of 29 investigated 2.2. User-music bipartite networks
age groups our sample had over 5,000 users, and in seven Since each FP contains the favorite songs of a user, these
groups over 30,000 individuals, with the maximum number FPs allow us to construct and measure bipartite networks be-
at the age of 21 (a total of 49,870 users). tween users and songs [48, 66]. In the bipartite network,
Given that the year 1990 is set as the default year in NCM each user corresponds to an FP and connects to every track
apps, we observe in Fig. 1(b) an anomalous peak at the age in the generated FP, as shown in Figure 2.
26. For the remainig user ages, we can see that the number of A bipartite network consists of nodes that can be divided
users increases as they are getting younger, peaking around into two disjoint and independent node sets, such that every
the age of 21. network link connects from one node set to the other (see
Fig. 2). For example, a user-content bipartite network model
Table 1: The tag-set of NCM
has been employed recently to measure users’ influence and
Tag Class Tags
its diffusion in online social communities such as Pinterest
Language Chinese, EU&US, Japanese, Korean, Cantonese, Other LNGs [66]. In our case, in the user-music bipartite network of
Pop, Rock, Folk, Electronica, Dance, Rap, Light Music, Jazz, Country, NCM, each user is connected with each track listed in her
R&B/Soul, Classical, Ethnic, Britpop, Metal, Punk, Blues, Reggae,
Genre or his FP, as depicted in Figure 2. We describe the global
World Music, Latin, Alternative/Indie, New Age, Antique, Post-Rock,
Bossa Nova features of user-music bipartite networks based on three pa-
Early Morning, Night, Studying, Working, Noon Recess, Afternoon
Scenario rameters: 1) the length of user’s FP corresponding to how
Tea, Metro, Driving, Sports, Travel, Walking, Bar
Nostalgia, Refreshing, Romantic, Sexy, Sad, Healing, Relaxing, many songs a user saves in the FP, 2) the number of users for
Emotion
Lonely, Touched, Exciting, Happy, Quiet, Missing
Theme
OST, ACG, Campus, Game, 1970s, 1980s, 1990s, Web Song, KTV, each track, and 3) the total amount of preferences from all
Classic, Cover, Guitar, Piano, Instrumental, Children, Ranklist, 2000s users for each respective track.
To estimate user’s musical preferences, we assume that
As shown in the inset of Fig. 1(b), the sex ratio of users each track in a user’s FP shares the same preference weight,
varies by age, with the ratio of male-to-female users increas- and each user’s preference has equal weight as well. Thus,
ing significantly with age. This ratio increases faster among when considering the total amount of preferences that a user
4
 as musical genre information. However, besides the tags of
(a) (b)
General Playlists that are assigned by users, NCM does not
provide any further information on a track’s category. To ob-
Slope = -1.43
tain the tag information for each track, we map in our analy-
sis the playlist’s tags to all the tracks in the playlist, and we
proceed with the following tag mapping steps:

P(A)
P(L)

i) According to the full tag-set of NCM, to show the initial-

Slope = -1.53
ization of track’s tags, we assign to each track in the dataset
a vector-set that contains five zero vectors, where each vec-
P(K)

tor corresponds to a tag class, and the number of elements in

each vector is equal to the number of tags in its correspond-
K
ing tag class.
ii) For each track, if a playlist with tag information (e.g.
L A tags A, B, and C) contains a given track, we add a unit value
Figure 3: (Color online) Three distributions of the user-music bipartite net- to the track’s vector elements that corresponds to tags A, B,
work. (a) The distribution of the length of user’s FP. The inset in this figure and C.
shows the distribution of the number of followers of each track. (b) The iii) After applying the above steps to all the tracks and
distribution of users’ attention focused on each track.
all the playlists, we normalize the elements in each non-zero
vector in each track’s vector-set. For example, if a track’s
has over L pieces of music in the FP, the amount of the user’s vector of the tag class “Language” is [5, 2, 2, 1, 0, 0], its nor-
preference for each individual track is 1/L. By summing up malized vector is [0.5, 0.2, 0.2, 0.1, 0, 0]. Other zero vectors
all users’ preferences for a specific track across all different remain unchanged.
FPs, we can calculate the total amount of users’ preferences A track’s tag information can be shown in its correspond-
or the total amount of attention given to one particular track. ing vector-set. For example, for a track with the vector
For example, two users can save the track a in their FP (see [0.5, 0.2, 0.2, 0.1, 0, 0] of the tag class “Language”, and with
Figure 2). One user has three tracks in her FP, so that track each element in the vector respectively corresponding to
a receives 1/3 unit of preference from that user; another user the tags “Chinese, EU & US, Japanese, Korean, Cantonese,
has two tracks in the FP, thus the same track receives 1/2 Other Languages”, it would then mean from this particular
unit of preference from this other user as well. In total, the tag order that the track would be highly related to Chinese
resulting total preference for the track a, or the total attention music and also slightly related to the music of the EU & US
given to it, is A1 = 5/6 (i.e. 1/3+1/2) unit (see the top right areas as well as the Japanese music, but a much weaker re-
track in Fig. 2). lationship with Korean or Cantonese music or the music in
Based on this bipartite network approach and the above other languages. In our following discussions, each value in
calculation of users’ attention, the global features of the bi- the vector-set is called the tag’s strength of the track. Apply-
partite network and the patterns of users’ attention can be ing this mapping method to all 4,261,266 tracks, we obtained
described by the following three distinct distributions: The the tag information for a total of 1,464,448 tracks (34.37%).
distribution of the length of a user’s FP, the distribution of Similarly, we can further map tracks’ tags to the FPs of ac-
the number of followers of each track, and the distribution tive users. We also assign each FP a zero vector-set, and add
of users’ attention A given to each track. Figure 3 shows the the tag vector-sets of every track in the FP to the FP’s vector-
distribution patterns of these three global features of the user- set, and we then normalize the elements in each non-zero
music bipartite network. First, we see that the patterns of all vector in the FP’s vector-set. The FP’s vector-set also reflects
three distributions are strongly heterogeneous. The distribu- the user’s tag information. This way, we were able to obtain
tion of the number of NCM users per song and the distribu- the tag information of 3,021,208 active users (99.76%).
tion of total users’ preferences per song, as shown in Figure Furthermore, in our analysis that divided users into dis-
3, are well fitted by power-law-like functions. These hetero- tinct groups, we applied a similar tag mapping method that
geneous properties have been observed widely as a hallmark adds up the tag vector-set of each user in the group and nor-
of human behavior across many different online social me- malizes the cumulative vector-set to represent the vector-set
dia, indicating that only a few popular songs typically attract of the group. We also obtained the global vector-set that cu-
the attention of most users, while a relatively small number mulates and normalizes the vector-sets of all users. Figure
of active users are actually the carriers of the majority of 4 shows the tags’ strength of the global vector-set for male
musical preferences. Interestingly, from the curve damps at and female users. Obviously, the distribution of tag strength
L = 50, 100, and 1000, shown in Figure 3(a), we can observe in the tag classes of “Genre” is more heterogeneous, with
the impact of specific rules set by NCM such as the limits on a few genres, such as “Pop”,“Folk ” and “Electronica” at-
the maximum length of FPs that were set to 50 and 100 in tracting most users’ preferences. Several tags, as denoted
the previous NCM versions, and to 1000 in the later NCM by asterisks in Figure 4, are associated with large gender
version. differences. For example, female preferences for the tags
“Korean”, “OST” and “Romantic”, are obviously higher than
2.3. Tag mapping those of male users, while “Metal”, “Electronica”, “Game”,
In our analyses related to users’ musical preferences, it and “Exciting” tags show the opposite pattern of gender dif-
was necessary to obtain music category information, such ferences, reflecting the gender divide on musical tastes. In
5
 Number of tracks

Community Size
Rank of Community Size

Figure 5: (Color online) The distribution of the NCM community size in

Male
the Zipf’s ranking plot. The communities are ranked by the total number of
music tracks and users. The blue and red data points, respectively, denote
the numbers of tracks and users in each community.

same community of the user-music bipartite network. There-

fore, by investigating the community structure of songs in
the bipartite network, we can measure the diversity of mu-
sical genres that are typically preferred by the NCM users.
Figure 4: (Color online) The global tag strength distribution as a function It is important to note here that the identification using com-
of tag class. The asterisk symbols indicate the tags with large gender dif- munity structure is completely socially-determined, because
ference, with pink asterisks showing the tags for which female users have a even though two types of music may have somewhat dif-
higher strength than males, and correspondingly, the blue asterisks indicate
the tags with greater strength for male relative to female users. The num-
ferent style characteristics they would still be identified as
bers displayed next to asterisk symbols represent the tag-strength vs. gender belonging to the same community if they share similar lis-
ratios. teners.
We detect the listenership communities in our user-music
NCM network by employing the fast-unfolding algorithm in-
the following analyses, we only focus on FPs, whereas the troduced by Blondel et al. [72]. This algorithm retains both
information on General Playlists is not used anymore. the higher accuracy and the lower algorithmic complexity,
and is widely used in network-based analyses of social media
2.4. Community detection [73]. Without differentiating each side of the nodes (either
music or users) in this bipartite network, the fast-unfolding
Although the tag mapping method can reveal a lot about algorithm would identify a community that includes at least
prevalent music genres and corresponding users’ prefer- one user and one song that was included in the user’s FP.
ences, tag mapping cannot capture all information on the In total, there were 1656 initially detected communities in
NCM’s tracks and users. Since we have established a user- this user-music bipartite network. As shown in Figure 5, the
music bipartite network, community analysis, as a network- size distribution of these detected communities is highly het-
based clustering-like method [67], can serve as an efficient erogeneous, such that 99.47% of users and 96.53% of tracks
approach to exploring tracks’ categories and user’s musi- can be grouped into eight largest communities. Our Table 2
cal preferences. However, unlike unipartite networks which shows the information related to these eight largest commu-
have been investigated extensively in past years with a vari- nities.
ety of community identification algorithms [68, 69, 70], the In addition, using the above tag mapping method, we
methods for community detection in bipartite networks have calculated the tag vector-set of the identified eight largest
been less frequently studied and applied to real-world net- communities. For each community, the tag with the largest
worked systems [71]. strength in each tag class is called the first primary tag, and
In a typical complex network, some nodes can be identi- the second largest one is the second primary tag. As shown
fied as a group if they are densely connected with each other in Table 2, the first and the second primary tags of each com-
while their connections outside of the group are relatively munity are very different, possibly reflecting the individual
sparse. This type of node grouping in a network is called a differences in musical preferences, the underlying moods, or
“community”, and we say that groups of nodes form a com- even the wellbeing status of users in different communities.
munity structure within the network. Generally speaking,
a community structure can be identified if network nodes
2.5. Diversity analysis
can be grouped into sets such that each set of nodes ex-
hibits dense internal connections [67]. For example, users Diversity analysis has been used widely in the study of
with similar musical tastes are more likely to save songs musical preferences [32]. Following Eagle et al. [74], and
from same or similar genres in their FPs; thus, songs from based on the tag information and the communities detected
the same or related genres would typically appear within the in our dataset, we define a series of metrics to describe and
6
Table 2: The eight largest communities of NCM. The communities are ranked by size, ranging from large to small. ρT and ρU , respectively, are the proportion
of tracks in global total tracks, and the proportion of users in global total users of each community. The rightmost five columns show the communities’ first
and second primary tags in each of the five tag classes.
Commuity ρT ρU Average age Female proportion Language Genre Scenario Emotion Theme
A 0.0819 0.4006 22.7 0.337 Chinese, Cantonese Pop, Folk Night, Working Nostalgia, Sad 1990s, Classic
B 0.1582 0.2297 21.7 0.289 EU&US, Chinese Electronica, Pop Sports, Night Exciting, Relaxing Game, 1990s
C 0.3615 0.0591 24.8 0.299 EU&US, Chinese Light Music, Rock Night, Studying Quiet, Relaxing OST, Classic
D 0.1903 0.1048 20.5 0.331 Japanese, Chinese Light Music, Antique Night, Studying Healing, Relaxing ACG, Game
E 0.0430 0.0729 20.7 0.667 Korean, EU&US Pop, Rap Night, Sports Relaxing, Healing 1990s, OST
F 0.0719 0.0458 28.7 0.247 Chinese, EU&US Pop, Ethnic Night, Driving Nostalgia, Quiet Classic, OST
G 0.0314 0.0569 22.0 0.352 Chinese, EU&US Folk, Rock Night, Travel Quiet, Lonely Guitar, Classic
H 0.0271 0.0249 23.0 0.423 EU&US, Other LNGs Pop, Ethnic Driving, Night Relaxing, Exciting Classic, 1990s

assess various types of diversity of musical preference be- a tag class G as:
haviors. For the i-th user, given her normalized tag vector
(υi1 , υi2 , ...) of a given tag class (e.g. tag class G), her indi- DGT U = DGT A − hDGT I i. (3)
TI
vidual tag diversity DGi of the tag class G is then defined as
It measures the difference among users within the same de-
[47]:
P mographic group on a given tag class.
− j υi j log(υi j )
TI
DGi = , (1) Similarly, we can define three types of community diversi-
log mG ties based upon the detected communities in the investigated
where υi j is the tag strength of the j-th tag of tag class G, NCM network. For the i-th user, her individual community
and mG is the total number of tags in the tag class G; for diversity DCI
i is given by:
example, mG = 6 for the tag class “Language”. The numer- P
− j vi j log(vi j )
ator of the expression is actually the information entropy of DCI
i = , (4)
log Mc
the distribution of each element in the tag vector, and this
diversity measure addresses the degree of homogeneity of a where vi j is the proportion of the j-th community’s tracks
user’s preference for a given music tag in a given tag class. in the i-th user’s FP, and Mc is the global, total number of
For example, in the tag class G, if the user listens only to communities. DCI i equals to 0 if the user only follows the
the music with a single tag, then there is only one tag with tracks within the same community, and DCI i equals to 1 if the
a non-zero strength in the user’s tag vector of the tag class user’s FP tracks distribute homogeneously across each of the
TI
G, and DGi equals the minimum value of 0, indicating that existing communities. Thus, unlike individual tag diversity,
the user’s preference for the tag class is quite simple; more- individual community diversity measures the degree of ho-
TI
over, if each tag has the same tag strength, DGi equals the mogeneity with which the user’s FP tracks distribute across
maximum value 1, indicating that the user has equal prefer- different communities. The higher the value, the more ho-
ences for the music of each tag in the tag class. For a given mogeneous is the distribution of FP tracks across commu-
demographic group, the average value of hDGT I i of tag class nities; correspondingly, the lower the value, the more con-
G describes the average level of the tag diversity across all centrated are the user’s FP tracks within only a few commu-
users’ individual preferences in that group. nities, representing thus the less diversified preference. The
Similarly, for a given demographic group of users, obtain- value of hDCI i describes the community diversity when av-
ing the group’s normalized tag vector (gG1 , gG2 , ...) of the tag eraged across all users’ individual preferences in the group.
class G, the aggregated tag diversity DGT A of the tag class G For a certain demographic group, we further define the
is defined as: aggregated community diversity as:
P
− j q j log(q j )
P
− j gG j log(gG j )
TA
DG = , (2) DCA = , (5)
log mG log Mc

where gG j is the tag strength of the j-th tag in the tag class where q j is the proportion of the j-th community’s tracks in
G of the group’s tag vectors. It describes the tag diversity all FPs of the users in the group:
of the group’s users in an aggregated fashion, namely, DT A P
k lk j
treats users in a certain demographic group as a whole. qj = P , (6)
k Lk
In order to analyze the differences among users within the
same demographic group, we need to understand the rela- where lk j is the total number of the j-th community’s tracks
tionship between individual diversity DT I and the aggregate in the k-th user’s FP, Lk is the total number of tracks in the k-
diversity DT A . If all users in a certain demographic group
P
th user’s FP, and k denotes the summation over all users in
like tracks with similar tags, the value of DT A will be close the group. This metric describes the aggregated diversity of
to DT I . If all users’ preferences for tags in the same group are users’ preferences for tracks in a certain demographic group.
different, the gap between the DT I and DT A would be large. A higher value indicates a more diversified taste in music of
Therefore, the aggregated tag diversity DT A of a certain de- a certain group of users.
mographic group contains two components: The individual Similar to the case of tag diversities, the aggregated com-
tag diversity of each user, and the within-group diversity of munity diversity DCA of a certain demographic group also
users. We thus define the within-group diversity of users for contains the individual tag diversity of each user and the
7
within-group diversity of users. We therefore use the dif- ability to attract people’s attention. For example, the influ-
ference: ence of a novel cultural product usually attains its maximum
DCU = DCA − hDCI i, (7) peak after a diffusion process, and then trends to a long-term
to measure the within-group community diversity of users. decay. We expected that the evolution of music popularity
would show similar features: Given the potentials of mod-
2.6. Group Difference Analysis ern online social media with rapid information dissemina-
The Kullback-Leibler divergence (KLD) is an extensively tion, the term of the diffusion process of a novel musical
employed measure for a difference between two probability product before the peak should generally be shorter than its
distributions [75]. Generally, the KLD from a normalized subsequent decay stage. With this consideration in mind, our
discrete distribution (P) to the reference normalized discrete analysis of the evolution of music popularity focuses on the
P P(i)
distribution (Q) is: ∆(P||Q) = i P(i) ln Q(i) for each i sat- mode of the decay term of music influence.
isfying P(i) > 0 and Q(i) > 0. The value of KLD is non-
negative. A KLD of 0 indicates that the two normalized
distributions are exactly the same. A higher KLD means a
larger difference between the two distributions. Its value of
1 means that the two distributions are so different that we
All
cannot approach one of them from the other. Since ∆(P||Q) Female
Male
generally does not equal ∆(Q||P), the form employed in our
analysis is the mean value of ∆(P||Q) and ∆(Q||P):
1
∆= [∆(P||Q) + ∆(Q||P)]
2
1 X P(i) X

Q(i)  (8)
=  P(i) ln + Q(i) ln .
2 i
Q(i) i
P(i) 

Our group difference analysis mainly focuses on gender

differences in users’ musical preferences. Similar to the case
of the diversity analysis, gender differences have two types
of representations: One is the difference between the vector-
sets of tags of the two groups, and the other is the difference
between the community distributions of users’ FP tracks in
the two groups. The KLD, therefore, has two different defi-
nition levels in our analysis: One is tag-based and the other
one is based upon the detected communities in our investi-
gated NCM dataset.
For each studied demographic group, we separate male Figure 6: (Color online) The variation of the average global preference IG
and female users into two subgroups. For the male user sub- of tracks as a function of the time interval between the release year of the
group with its normalized tag vector (g1M , g2M , ...) of the tag music track and the data collection year (2016) in a log-log plot. Panel
(a) shows the associated gender differences; the inset depicts the same data
class G, and for the female user subgroup with its normal-
in a semi-log plot. Panel (b) shows the decay mode for the tracks with
ized tag vector (gFf1 , g2F , ...) of the tag class G, the associated different attention levels. The solid line in each panel represents the fitting
tag-based KLD measure for gender differences within the tag function for all tracks, which is power-law-like with the exponential tail
class G is defined as: y = 1.97x−0.34 exp(−0.023x). All tracks published in the same year were
  ranked in accordance with their total attention received; the top 5% of tracks
T 1 X M giM X F giF  in this attention ranking are the ”hot tracks”, the tracks ranked between the
∆G =  gi ln F + gi ln M  . (9) top 5% and top 25% are the ”middle tracks”, and the rest are the ”unpopular
2 i i g i ig tracks”.

Similarly, for a given demographic group, obtaining the In the NCM dataset, each track belongs to an album, and
normalized proportion qiM of the i-th community’s tracks in
the dataset includes the release year of each album. The
all users’ FPs in the male subgroup, and the corresponding
dataset also includes users’ birth dates. Similar to the method
normalized proportion qiF for the subgroup of female users,
used to calculate a track’s total attention A, we now define a
the community-based KLD for gender differences can be cal-
matrix AY to describe the attention map between users with
culated by: different birth years and tracks with different release years.
Each element of the matrix AY is then:
 
C 1 X M qiM X F qiF 
∆ =  qi ln F + qi ln M  . (10)
2 i qi qi 1 X X εi j
i AYij = (11)
n j N L Lk
k
3. Results and discussions
where i and j denote the i-th release year of a track and the
3.1. The decay of music influence j-th birth year of a user, N is the total number of users in the
Most cultural products are typically characterized by their dataset, Lk is the number of tracks in the k-th user’s FP, and n j
life cycle and the associated diffusion pattern related to their is the total number of users born in the j-th year. εi j = 1 if the
8
 and female users (Fig. 6(a)), nor did they show significant
differences among tracks with varying popularity (Fig. 6(b)).
Remarkably, similar slow-falling patterns with a power-
law-like tail have recently been identified in the temporal
variation of different online behaviors, e.g. in the time se-
ries characterizing the outbreak of internet memes or in the
spreading patterns of popular topics [63, 64]. However, a
noticeable difference is that the decay of music preferences
that we observed in the NCM dataset is much slower than in
any other case previously observed in online social media,
which is in part corroborated by a more recent study on the
Male online attention given to cultural products such as movies,
biographies, and popular songs [76].
The value of the characteristic exponent of the power-law
Figure 7: (Color online) The release year distribution of user’s FP tracks function for musical preferences in our study is 0.34, which
for the users with a different birth year. Panel (a) shows the global pattern, is also lower than what was typically reported for other cases
where the dotted-line stands for the median of the distribution. Panel (b) in online social media. The exponent tail denoting the stage
shows the corresponding patterns of female and male users, and the dotted-
lines depict the medians of each distribution. The shadowed areas (light blue of rapid decay often emerges when the time period is longer
shadow for male users, light pink shadow for female users, and the light pur- than 20 years, which closely corresponds to one generation,
ple shadow for their overlap) show the regions between the corresponding thus implying that the rapidly decaying cultural taste is a
upper and lower quartiles.
generational phenomenon. Within the same generation, the
cultural preferences are much more stable.
track in the FP of the k-th user was released in the i-th year
and the k-th user was born in the j-th year; otherwise εi j = 3.2. The relationship between users’ age and sensitivity to
0. Obviously, each column of AY is actually a normalized music
distribution of the release year of the favorite tracks for the Age and gender have often been identified as important
users with the same birth year. factors in shaping musical attitudes and preferences [28, 38].
Similarly, the global distribution of the release years of A common support for this influence of demographic factors
favorite tracks of all users can be obtained as: on preference behavior is e.g. the observation that teenagers
typically pay far more attention than adults to new pop music
1 X X εi
AGi = , (12) trends [28].
N N L Lk To further analyze age trends in the context of Chinese
k
users’ musical preferences, we calculated the median, 25%,
where N is the total number of users, and εi = 1 if the track and 75% quartiles of their preferences over the songs that
in the FP of the k-th user is released in the i-th year, and oth- were released across different years spanning the period of
erwise εi = 0. The global distribution describes the average several decades. Figure 7(a) shows the central region of the
level of users’ preferences for the tracks that were released area between the upper and lower quartiles of preferences
in different years. by the age cohort. For older listeners, we can see that they
Since the total number of tracks released in different years clearly prefer older songs covering a wider range of release
can vary, we calculate the average global preference as: years than younger listeners. On the other hand, younger
listeners evidently pay more attention to more recently re-
AGi 1 X X εi leased tracks. Figure 7(b) depicts the associated gender dif-
IiG = = , (13)
ρi ρi N N L Lk ferences. Even though they both display a similar trend, es-
k
pecially around and before the age of 20, women are more
where ρi is the ratio of the number of tracks released in the likely to prefer more popular and newly-released songs than
i-th year, namely, ρi = mi /M, where mi is the total number men, suggesting that female users are more sensitive than
of tracks released in the i-th year, and M is the total number men to the influence of musical trends. This finding is in line
of tracks. Generally, the average preference for old tracks both with earlier offline studies of musical preferences [36]
is much lower than for recently released tracks. Naturally, as well as with the more recent analyses of large-scale music-
older users show higher preferences for older tracks. based online communities such as Spotify [77]. To control
We next addressed the relationship between users’ age and for cross-cultural effects, the study in Ref. [77] investigated
the release year of their preferred music tracks, as well as only US Spotify users. Excitingly, this remarkable similar-
the question of how fast does a user’s taste in music decay ity between our findings for NCM Chinese music listeners
over time. In Fig. 6, we re-plot the curve of I G by rescal- and the results for US Spotify users suggests the existence
ing the x-axis to the difference between 2016 and the track of cross-cultural universals in preferences for popular music
release year. Fig. 6 thus shows the decay of users’ musical that hold across a wide variety of demographic factors.
preferences: The pattern is well fitted by a power-law-like As a further part of our analysis, we investigated the mu-
function with the exponent tail y = 1.97x−0.34 exp(−0.023x). sical sensitivity of NCM users. We defined the matrix AY
Such power-law-like decaying patterns for musical prefer- shown in Eq. (11) to describe the distribution of users’ at-
ences did not reveal a significant difference between male tention to the tracks released in year i for the users with birth
9
 Figure 8: (Color online) The heat maps of the matrix ln(R): (a) the global pattern, (b) the pattern of male users, and (c) the pattern of female users.

year j, and the global attention distribution AGi . Here, we music reaches its peak around the age of 13. Thus, people’s
treat AG as the background distribution of user’s attention, sensitivity to music is the greatest when they are in their teen
and we use the matrix R of the ratio between AYij and AGi to ages. This finding resonates well with the common experi-
express the relative attention of users with a birth year j: ence that people tend to be most strongly influenced by the
music that has been popular in their youth. Moreover, this
AYij finding is corroborated by a recent big data analysis of US
Ri j = , (14)
AGi Spotify users [61], reporting the strongest influence on adult
musical preferences at the ages of 13 (for women) and 14
representing thus the attention of users relative to the global (for men).
average level. The maps of the matrix R for all users and This striking similarity between our findings and those ob-
the two corresponding cases for male and female users are served with US Spotify listeners highlights again the univer-
shown in Figure 8. We see again that users’ preferences gen- sal nature of musical preferences that holds across cultures,
erally depend upon their age: Older users pay considerably this time with respect to the period of the formation of adult
more attention to older tracks. musical tastes. We further observed that the fitting function
For the case i > j, the value i − j actually denotes the age is not symmetrical, and the decay of sensitivity after xc oc-
of users when the tracks were released. We therefore define curs faster than its growth before xc (w1 > w2 ), indicating a
the sensitivity S j (i− j) of users with birth year j for the tracks rapid decline in sensitivity to new music after adolescence.
released in year i, as the normalized value of R for each j:
We further observed an interesting relationship between
AYij * AYk j +
−1 gender and users’ musical sensitivity. As shown in Figure
S j (i − j) = Ri j hRk j i−1 = , (15) 9(b), the fitting curves for the average sensitivity of men and
AGi AGk women are different. The estimated values of the parame-
where k denotes all the release years of tracks, and angle ters y0 , xc , H, w1 and w2 , respectively, are 0.41, 12.76, 0.85,
brackets represent the operator for the average value. Here 14.25 and 7.19 for male users, and 0.50, 13.79, 0.83, 12.83
we set that S = 1, corresponding to the case where the users’ and 7.60 for female users.
attention is equal to the global average value of attention. On average, men thus reached their peak of sensitivity to
Figure 9(a) shows the users’ sensitivity to music as a func- music at a younger age than women; however, the estimated
tion of users’ age when the tracks were released, whereby peak age xc of male users was only slightly smaller than
negative numbers correspond to the track releases before the that of their female counterparts. Moreover, after passing
users’ birth year. The average sensitivity of users of each age the peak-age of sensitivity to music, women remained more
is depicted by the open blue circles which exhibit a unimodal sensitive than men at the same age to the most recent music
shape that can be nicely approximated by a Bigaussian func- trends. These results are further in line with what is revealed
tion: in Figure 7(b), namely, that adult female users are usually
 more sensitive to newly released music, while male listeners
x−x
y0 + H exp [−0.5( w1 c )2 ], x < xc

 develop an earlier and stronger ’taste freeze’ phenomenon.
y(y0 , xc , H, w1 , w2 ) = 
y0 + H exp [−0.5( x−xc )2 ], x ≥ xc
 Together, these findings are partly corroborated by the
w2
(16) study of US Spotify users [77], showing that men’s prefer-
where the estimated values of the parameters y0 , xc , H, w1 ences for popular music decrease faster than those of their
and w2 , respectively, are 0.43, 12.88, 0.87,13.18 and 7.26. female counterparts, lending thus support for the existence
The peak of the fitting curve thus corresponds to the value of a further cross-cultural universality in music taste forma-
xc = 12.88, meaning that on average, the user’s sensitivity to tion. However, a unique, culture-specific Chinese feature in
10
 for age, whereas “−−” means that the tag’s strength remains
largely unchanged across the three stages.
The age modes of tags for each tag class are listed in Ta-
ble 3. Generally, age modes with upward trends ( e.g. //,
/− and −/) indicate that a tag is more likely to attract older
users, whereas downward-trend age modes (e.g. \−, −\,
and \\) mean that a tag would attract more younger users.
These age modes of tags reveal rich information about mu-
sical age trends, cultural influences, users’ situational con-
texts, and pervading emotions in different age groups. There
are several examples: Some new-rising genres or themes,
such as “Electronica”, “Game”, or “ACG” have a decaying
age mode, suggesting that these music types mainly attract
young listeners. In contrast, traditional or sophisticated mu-
sic genres (e.g. “Classical”, “Ethnic” and “Jazz”) mainly
show growing age modes, which is generally in agreement
with findings reported in Ref. [38]. The decaying age
modes of tags “Japanese” and “Korean” indicate the rapidly-
growing cultural influences from Japan and South Korea on
Male
young Chinese people. The age modes observed for some
tags in the class “Scenario”, e.g. the growing mode of “Driv-
ing” and “Travel”, and the decaying mode of “Studying” and
Figure 9: (Color online) (a) The sensitivity of NCM users to the tracks “Sports”, show differences in the situational contexts among
released at their different ages, whereby negative age represents the cases users with different ages. In the tag class “Emotion”, the
in which tracks were released before the birth of the user. The blue open tags that are related to melancholic moods, e.g. “Nostalgia”,
circles are the average values of all data points (gray dots) of sensitivity, and
the black line shows the fitting Bigaussian function. (b) Gender difference
“Lonely” and “Missing”, usually have a growing mode. This
for the averaged sensitivity data as a function of users’ age when the tracks suggests a tendency for melancholic or negative emotions to
were released. The dark blue and dark red lines respectively are the fitting grow with age among Chinese music listeners.
Bigaussian functions for the cases of male and female users. We further investigated the relationship between age and
tag diversity. For each tag class, the average individual
the investigated data could be the observed earlier peak for tag diversity hDT I i, the aggregated tag diversity DT A , and
men than for women, since US music listeners showed either within-group tag diversity DT U of male and female users
roughly equal patterns for both male and female listeners in in each age group, are plotted in Figure 10(a), (b) and (c),
their teens [77], or the reverse pattern, with women showing respectively. There are significant differences between the
an earlier music sensitivity peak than men [61]. However, diversity levels of various tag classes. The tag class “Lan-
this uniqueness of Chinese listeners in our analysis may also guage” has the lowest hDT I i (Figure 10(a)) and the highest
be due to limitations in the available dataset, such as the gen- DT U (Figure 10(c)) for the majority of age ranges, indicat-
erally larger male subpopulation that was present in our stud- ing a strong individual heterogeneity and the largest within-
ied sample of NCM users. group diversity of users’ preferences for the language of mu-
sic tracks, suggesting that a typical NCM user prefers to lis-
ten the music that is related to just a few languages, and that
3.3. Genre-related age trends and gender differences the language preferences within an age group are usually di-
Besides discovered general age trends for popular mu- versified. The tag class “Genre” also has a lower hDT I i, the
sic, age and gender as demographic factors may also exhibit lowest DT A , and its DT U is also in the lower range (Figure
genre-specific influences on musical preference behavior. To 10(a-c)), suggesting small within-group differences in users’
explore the genre-related age trends in users’ musical prefer- preferences for musical genre. On the other hand, the tag
ences, we firstly calculated the tag-vectors of each age group, classes “Emotion” and “Scenario” have much higher hDT I i
and we then divided the age ranges from 12 to 40 into three and DT A , and the lowest DT U .
separate age stages: 12-18, 19-25, and 26-40, and finally, Furthermore, some tag classes show conspicuous gender
we averaged the tag strength within each age stage for each differences. As shown in Figure 10(a), male users have
tag. Furthermore, for each tag, we compared the average higher hDT I i for “Genre”, possibly indicating that men have
tag strengths for the pairs of two continuous age stages. If a greater openness to a variety of music preferences, and
the latter is higher than the former 10%, we used the sym- lower hDT I i on “Emotion”, relating to the possibility that
bol “/” to express the change trend between the two age women are more likely to respond to music in a more emo-
stages. We further used “\” if the latter is lower than the tional way than men. Female users also have a higher within-
former 10%, and “−” for the rest of the cases. This way, group diversity DT U for the tag class “Language” (Figure
we can obtain the age modes of each tag, which is expressed 10(c)), especially for users under the age of 25.
by combining two symbols from “/”, “−” and “\”. For ex- In all of the three types of tag diversities, there is a remark-
ample, “//” means that the strength of the tag continuously able transition around the age of 25, including the downward
increases with age, and “/\” denotes a bell-shape-like trend transition on hDT I i for all tag classes (Figure 10(a)) and DT A
11
 Table 3: The mode of age trends of each tag for male and female users
Gender Tag Class // /− −/ −− \/ \− −\ \\
Japanese,
Language Other LNGs Cantonese Chinese EU&US
Korean
Folk, Rock, Country, World Music,
Ethnic, Jazz, Pop, Light Music, Antique, Dance, Electronica,
Genre Britpop, Post-Rock, Classical, Metal, R&B/Soul
Blues, Punk, Latin Alternative/Indie Reggae Rap
Bossa Nova New Age
Night,Walking, Afternoon Tea,
Driving,
Female Scenario Travel Early Morning, Bar, Noon Sports, Studying
Working
Recess, Metro
Relaxing, Healing, Quiet, Sad,
Emotion Nostalgia Lonely, Missing Refreshing, Happy Exciting
Touched, Romantic, Sexy
Classic, 1980s,
Web 1990s, Game,
Theme Instrumental, Guitar, KTV Children OST, Campus, Piano Cover
Song Ranklist ACG, 2000s
1970s
Language Chinese Cantonese Other LNGs EU&US Korean Japanese
Rock, Ethnic, Pop, Antique, R&B/Soul, World Light Music,
Electronica,
Genre Jazz, Blues, Latin Folk, Country Classical, New Music, Britpop, Post-Rock, Metal Alternative/Indie,
Rap
Age, Punk Bossa Nova Dance, Reggae
Night, Travel, Walking, Early Morning,
Male Scenario Studying, Bar Sports
Afternoon Tea, Metro Driving, Working, Noon Recess
Quiet, Sad, Lonely, Relaxing, Healing, Refreshing,
Emotion Nostalgia Happy Exciting
Missing Touched, Romantic, Sexy
Classic, Guitar, Instrumental, Game,
Theme Campus, KTV 1990s, Ranklist, Cover, Piano Web Song
1980s, 1970s OST, Children ACG, 2000s

None of the tags were observed to show the trend “/\”, which was thus not included in this table. The full description of trends is given in Section 3.3.

and DT U for the tag class “Language”(Figure 10(b),(c)), as can observe that this decline after the age of 25 is clearly
well as the upward transition in DT U for all tag classes but stronger for women than for men.
“Language” (Figure 10(c)). Under the age of 25, users can At the aggregate level, the transition at the age of 25 is
rather stably maintain a higher level of individual tag diver- more obvious. As shown in Figure 10(e), there is an ob-
sities hDT I i. After the age of 25, individual diversities tend vious steep rise starting around the age 25, indicating that
to decline, and besides the tag class “Language”, the within- the aggregate community diversity becomes increasingly el-
group diversities hDT U i for most of the tag classes start to evated. The gender influence appears to be weaker than that
rise. This transition suggests that users under the age of observed at the individual level, whereby female users have
25 usually have a higher openness towards different types a higher DCA in the ages between 14 and 30.
of music, and then develop stronger musical preferences for We further explore the within-group community diversity
some particular music types with a growing preference bias (DCU = DCA −hDCI i) of male and female users from different
among different users. The tag class “Language”, the only age groups. As shown in Figure 10(f), there is a remarkable
exception to the growing trend of hDT U i after the age of 25, valley for the age groups between 20 and 25 for both men
is closely linked to the cultural influences from other coun- and women, suggesting that musical preferences of differ-
tries, and its declining pattern with age potentially reflects ent users are more homogeneous within this range of ages.
a stronger level of multiculturality present among younger There is also a steep rise between 25 and 28 years of age,
Chinese people. which is afterwards turning into a stable growth trend with
Given these results, the question that naturally arises is a rapid separation between male and female users. This dif-
how to interpret this transition seen around the age of 25? ference in DCU is for female users generally higher than the
The following analysis on community diversities provides one for their male counterparts (Figure 10(f)).
an insight into this question. As we have mentioned before, We notice in passing that the age group of 20-25 is most
different from the tag information, users’ music preferences likely composed of college students or recently-employed
cannot be reflected in the community structure if two types of college graduates, whereas the age group of 25-28 usually
tracks with a different style have almost the same followers, corresponds to the initial career-growth stage. The higher
and the patterns of users’ music preferences that can be ex- homogenization of musical tastes among college students
tracted from the community structure are the cases that have and young professionals could therefore be attributed to the
driven an obvious social division. In other words, some of more homogeneous social structure in these environments.
music tracks’ social attributes can be revealed by the detec- Thus, the transition at the age of 25 is more likely to re-
tion of their underlying community structure. flect the change of the social environment of an individual,
Specifically, we compare the three types of community from a campus-based university life to the life of the devel-
diversity measures that we mentioned previously for gen- oping career and the growing pressures to establish a family
der and age groups. As shown in Figure 10(d), the aver- [78]. This interpretation fits nicely with our previously dis-
age individual community diversity hDCI i across different cussed observations for the age modes of genre-related tag-
age groups forms an inverted U-shape-like curve for both ging behavior (see Table 3), showing that at this stage of life,
men and women. Here, the individual community diversity Chinese listeners typically develop preferences for unpreten-
increases prior to the age of 20, reaching its peak between tious and sophisticated styles of music, which is also in line
the ages 20 and 25, and then declines after the age of 25, with the findings of Bonneville-Roussy et al. [28]. More
suggesting that listeners become less active in exploring dif- specifically, unpretentious music styles (e.g. pop, country,
ferent types of music in their early adulthood. Moreover, we ethnic, folk) with their relaxing themes may particularly be
12
 (b)

DTA
Male

Figure 10: (Color online) Changes of the different types of diversity with age. Panels (a), (b), and (c), respectively, are the average individual tag diversity
hDT I i, the aggregate tag diversity DT A , and within-group tag diversity of users DT U , for gender and different tag classes. Panels (d), (e), and (f), respectively,
are the average individual community diversity hDCI i, the aggregate community diversity DCA , and within-group tag diversity of users DCU for different ages
and gender groups. The vertical dashed lines highlight the age of 25.

appealing to individuals with the family life focus, whereas The largest gender differences for various tag classes can be
the artistic and creative qualities of sophisticated music (clas- observed at the age of 13, which surprisingly coincides with
sical, jazz, blues etc.) are more attractive among adults fo- the age of the listeners’ highest musical sensitivity (see Fig.
cusing on their career and social status development [28]. 9). Prior to the age of 25, the tag class “Language” shows
a much larger ∆T relative to all other tag classes, suggesting
a stronger gender-specific sensitivity to various cultures at
a younger age. Besides these two KLD measures, we used
the Jensen-Shannon divergence (JSD) to additionally quan-
tify the underlying gender differences in users’ musical pref-
Theme
erences and their variability with age. The JSD measure is
C

actually the symmetrized version of KLD, and as such has

widely been used for the assessment of differences across
distributions. This additional analysis revealed that both the
tag-based and the community-based JSD measures resulted
Age
in almost identical trends as those obtained with the two
KLD measures that were depicted in Fig. 11 (not shown).
A summary of notations and symbols for different measures
used throughout this paper is shown in Table 4.

3.4. Relationship between music preferences and regional

economic development
Figure 11: (Color online) The tag-based KLD (∆T ) for gender differences
in each tag class as a function of age. The inset shows the community-based At both individual and societal levels, the previous re-
gender-differences KLD (∆C ) as a function of age. search has shown a strong connection between cultural pref-
erences and economic development [31]. Eagle and col-
Finally, tag-based and community-based gender differ- leagues [74] have discovered a remarkable association be-
ences in users’ musical preferences and their variability with tween the diversity of people’s social connections and the
age were also investigated in our present paper. We used regional economic development level of their location.
two KLD measures to quantify the underlying gender differ- Since users’ NCM information does not reveal their eco-
ences, including the tag-based KLD ∆T for each tag class, nomic status, we attempted to detect such an association at
and the community-based KLD ∆C . As shown in Figure 11, the societal level. From the yearbooks released by the Na-
both KLD measures show highly similar trends for all in- tional Bureau of Statistics of China and the provincial bu-
vestigated tag classes: They decay from a peak reached at a reaus of statistics, we collected regional economic indica-
teen age to a more stable level attained around the age of 25. tor data (including GDP per capita, per capita disposable in-
13
 Table 4: Notations and descriptions of main metrics used in the paper.
Notation Description Notes
The matrix describing the attention map between users with different birth Each column of AY is a normalized distribution of the release year of FP tracks
AY
years and tracks with different release years for the users with the same birth year.
It describes the global, average level of users’ preferences for the tracks released
AGi The global distribution of user’s preferences
in the i-th year.
The average global preference that tracks in the i-th release years received
IiG IiG = AGi /ρi , where ρi is the ratio of the number of tracks released in the i-th year.
from all users
The matrix describing the relative attention that tracks in the i-th release years
R Its element: Ri j = AYij /AGi
received from users with a birth year j
S j (i − j) User’s sensitivity to music with birth year j for the tracks released in year i S j (i − j) = Ri j hRk j i−1
TI The i-th user’s individual tag diversity within the tag class G in a certain It measures the degree of homogeneity of a user’s preference for a given music
DGi
demographic group tag in a given tag class.
TA
DG The aggregate tag diversity of the tag class G in a certain demographic group It measures tag diversity of a group’s users in an aggregate fashion.
The within-group tag diversity of the tag class G in a certain demographic DGT U = DGT A − hDGT I i, which measures the tag differences in a given tag class
DGT U
group among users belonging to the same demographic group.
The individual community diversity of the i-th user in a certain demographic It measures the degree of homogeneity with which user’s FP tracks are distributed
DCI
i group across different communities.
DCA The aggregate community diversity of a certain demographic group It measures the community diversity of a group’s users in an aggregate fashion.
DCU = DCA − hDCI i, which measures the diversity of user communities belonging
DCU The within-group community diversity of a certain demographic group
to the same demographic group.
It measures the gender differences in the attention distribution of users in a certain
∆GT The tag-based Kullback-Leibler divergence
demographic group for the tag class G.
C It measures the gender differences in the community distribution of user’s FP
∆ The community-based Kullback-Leibler divergence
tracks in a certain demographic group.

Figure 12: (Color online) The correlations between users’ music preferences and regional economic development. (a) and (b) The correlation between the
aggregated tag strength and the per capita disposable income at the province level for each tag. The tags are ranked along the decaying order of the correlation
for each tag class. (c) The correlation patterns for the tags with the strongest correlation in each tag class. The left panel shows the strongest positive
correlations, and the right one shows the strongest negative correlations. (d) The average correlation coefficients of each tag class for female and male users.
The coefficients for positive correlations and negative correlations are averaged, and exclude the tags with opposite correlations for male and female users.

14
come, and the output value of each type of industry) of all We further investigated the relationship between various
31 provinces (including all four municipalities and five au- tag diversity measures and regional economic development.
tonomous regions) in mainland China in 2015. Figure 13 shows the correlations between the three tag di-
Given that the population of China is dominantly Han Chi- versity measures and the per capita disposable income at
nese, and some provinces have a higher proportion of mi- the provincial level (Fig. 13(a)), and the typical patterns of
nority groups than others, the pertinent differences in musi- these relationships for the tag class “Language” (Fig. 13(b)).
cal preferences across various provinces could additionally Except for “Genre”, the average individual tag diversities
be affected by ethnicity-related factors, besides age, gender, (hDT I i) of all other tag classes are negatively correlated with
or economic development. We therefore excluded from fur- economic development. This may suggest that users in de-
ther analysis all provinces with an ethnic minority propor- veloped provinces are more likely to prefer only a few spe-
tion above 20% (including five autonomous regions: Inner cific kinds of music. On the other hand, the aggregated
Mongolia, Guangxi, Tibet, Ningxia, Xinjiang; and the three tag diversity (DT A ) and the within-group tag diversity (DT U )
provinces: Yunnan, Guizhou, Qinghai). We then focused our consistently show positive correlations with economic devel-
subsequent analyses on the rest of 23 provinces in mainland opment across different tag classes. These result patterns
China, that included a total of 249 different cities. may further suggest that economic development drives the
We treated users in each province as representatives of a diversification of musical preferences in the whole popula-
group and we first obtained the tag vector-set for each such tion, but not at the individual level. Moreover, similar to the
group. We then calculated the Pearson correlation coeffi- relationship between economic indicators and tag strengths,
cients between each tag’s strength and the indicator of eco- these correlations between various tag diversity measures
nomic development at the province level. Several economic and economic development are consistently higher for the
indicators, including GDP per capita, per capita disposable female than the male subpopulation of users.
income, and per capita tertiary industry output value, were However, the average individual community diversity
entered into the analysis. The per capita disposable income (hDCIi i) has the opposite patterns of relationship with re-
yielded the strongest correlations for most of the investi- gional economic development. For example, as shown in
gated tags. In our following discussions, we thus employ Figure 14(a), hDCI i seems to exhibit a weak positive correla-
per capita disposable income as the representative indicator tion with the per capita disposable income at the provincial
of economic development. level; however, this relationship was statistically not signif-
Figure 12(a) and (b) shows the correlations between each icant (p > 0.05). Actually, even a significant correlation
tag and the per capita disposable income. We see that the cal- would here not contradict the cases of the average individ-
culated correlations range from strongly positive to strongly ual tag diversity hDTi I i, but instead it would paint a picture
negative, indicating the existence of the relationship between in which users in developed provinces would have narrower
economic development and users’ preferences for some tags. musical preferences but wider links to other communities,
The cases with the strongest positive and strongest negative since the detected large communities are all tag-crossed.
correlations for each tag class are additionally plotted in Fig- The aggregate community diversity DCA had a signifi-
ure 12(c). It is noteworthy that, in the tag class “Emotion”, cant positive association (r = .60, p = .0027) with re-
most tags with positive emotional connotations also have gional economic development (Figure 14(b)), suggesting a
positive correlations with economic development, whereas rather strong relationship between economic development
tags with melancholic or negative emotional connotations, and users’ musical preferences at the aggregate level. The
such as “Lonely”, “Missing” and “Sad”, show negative cor- within-group diversity DCU , as measured by (DCA − hDCI i i),
relation with economic indicators. We note that this find- also positively correlated (r = .48, p = .021) with the re-
ing could also serve as a novel regional-level evidence for gional economic development (Figure 14(c)). Since the cor-
the relationship between economic indicators and individual relation between DCU and economic development is greater
well-being [79], as reflected via musically associated moods than that between hDCI i and economic indicators, it is easy
and emotions [1] in different regions in China. Moreover, for to find that the diversity of users seems to play here the ma-
the tag class “Language”, we find that the tag “Other LNGs” jor role and that economic development is mainly associated
(which stands for other languages), yields the strongest pos- with the growing diversification in the population of music
itive correlation with economic development. Thus, a higher listeners.
economic status can be associated with more opportunities In comparison to the province-level correlations, the re-
for the growth of niche cultures and subcultures in China. lationship between community diversity and regional eco-
When considering gender differences in our correlation nomic development at the city level was generally much
analyses, we found that on average, the tag strength of fe- weaker. Here, the measured correlations between regional
male users is stronger related to the level of economic devel- economic development and hDCI i, DCA , and DCU measures
opment than that of male listeners of NCM. In Figure 12(d), were r = .12 (p = .056), r = .22 (p = .00035), and r = .10
we see the average values of the Pearson correlation coeffi- (p = .102), respectively (see Figure 14(d), (e), and (f)). As
cients for each tag class split by gender (separately for posi- shown in Figure 14(d), the data spread around a Y-shaped
tive and negative correlations, respectively). We see that for region with several cities having low per capita disposable
the majority of tag classes, these correlations (both positive income but high average community diversity (hDCI i). The
and negative) are higher for female users than for males, in- patterns of DCU at the province level and DCA at the city
dicating that women’s musical preferences in China are more level in Figure 14(c) and (e) suggest a more intricate rela-
strongly tied to economic factors. tionship between economic development and musical prefer-
15
 (a) (b)
Language

<DTI>
r = -0.60

P = 0.002

Male

D TA
Language
DTA

r = 0.45

P = 0.031

D TU

Language
DTU

r = 0.60

P = 0.003

per Capita Disposable Income (in CNY)

Figure 13: (Color online) (a) The correlations between three types of tag diversities hDT I i, DT A , and (DT A − hDT I i) and the per capita disposable income at
the province level for each tag class. Gray, pink and blue bars, respectively, correspond to the cases of all users, female users, and male users. (b) The scatter
plots with the correlations between tag diversities hDT I i, DT A , and DT U of tag class “Language” and the per capita disposable income at the province level.
The currency unit of the per capita disposable income was CNY.

16
ences, that would thus require a further study across different information was assigned by the users themselves via their
societal levels and at various scales. tagging behavior. Such information is interesting because it
Similar correlations were also found between gender dif- contains individual perceptions of psychological and sonic
ferences and regional economic development. We calculated features of the considered musical genres. Thus, the ulti-
gender differences for users in each province using the tag- mate number of genres and their distribution across different
based KLD ∆T for different age groups. As shown in Figure tracks was in our study defined by the NCM users.
15(a) and (b), for all tag classes, the tag-based KLD ∆T cor- This approach is both innovative and relevant as there is
related negatively with the disposal income of each province. currently no general consensus about which and how many
Thus, the gender gap in users’ musical preferences decreased particular genre types should be studied when investigating
with an increasing provincial economic development. This musical preferences [27]. Besides, labeling music tracks by
finding suggests that economic development could thus con- a certain genre type can sometimes pose a nontrivial chal-
tribute to the reduced gender gaps in China when it comes lenge for music experts as well (e.g. if a given track con-
to development of musical tastes. Interestingly, this neg- tains a mixture of different genres), so that genre informa-
ative correlation between gender differences and income is tion originating from aggregate tagging behavior can yield
stronger in the older age group (26-40) than in younger age a more representative consensus about the various disputed
groups (Fig. 15(a)), even though the observed gender differ- cases. Of course, concomitantly, such behavior could also
ences generally tend to decrease with an increasing listeners’ introduce some uncertainty into the actual correspondence
age (Figure 11). In addition, we note that the outcomes of the between the assigned tags and the actual genre of a given
correlation analyses shown in Fig. 15 were almost identical track. However, it is rather difficult to estimate the magni-
when using the Jensen-Shannon divergence (JSD) instead of tude of this effect, given that quite often, only a few expert
the KLD measure (not shown). users can rather quickly and accurately set the tag of a track.
Similar to the study of Mellander et al. [31], we also found Furthermore, as pointed out by Bonneville-Roussy and
strong positive correlations between income and “sophisti- colleagues [28], earlier studies addressing the variability
cated” types of music (e.g. jazz, classical, world), but we of musical preferences with age employed only pop-music
did not observe such a relationship between economic fac- tracks (see e.g., [60]). Our present study instead analyzed
tors and what Mellander et al. [31] called the contemporary differences in preference behavior for a much wider range
music types (i.e. rap, soul, and reggae). This is an interesting of musical styles and situations in which music can be en-
cultural difference between US American and Chinese mu- joyed, allowing us thus to address the question of whether
sic listeners, suggesting that while we can still argue for the age trends in musical preferences vary by musical genre and
existence of cross-cultural universals in the development of context. While our genre sample is arguably sufficiently rep-
musical preferences, the relationship between economic fac- resentative of the Chinese musical tastes spectrum, we think
tors such as income and tastes for different musical genres is that even more diverse selections of music, including those
rather culture-dependent. Thus, various cultures may simply rated by different attributes than genre, should be used in fu-
differ in how they define “contemporary” music styles. For ture studies.
example, in the context of the social identity theory and its The limitations of user sample sizes and narrow age ranges
application to musical preferences development [80], peo- of music listeners that pervade earlier investigations of mu-
ple build their musical preferences jointly with others with sical preferences were overcome in our study by drawing
whom they share similar or same values, and these values a large sample of more than three million users distributed
are often reflected in shared interests in similar music types. across 29 different age groups (ranging from the age 12 to
So, obviously, values represented by American rap or reg- 40). For this range of listeners’ ages, on which our analy-
gae music are less shared among Chinese listeners, and as ses mostly focused, our sample included at least 846 users in
a result, different from what Mellander and colleagues [31] each individual age group. Remarkably, our sample had over
observed for their US American sample, we did not find any 5,000 users in 18 out of 29 investigated age groups, and in
correlations for these types of music in our NCM dataset seven groups we had over 30,000 individuals, with the max-
with Chinese music listeners. imum number at the age of 21 (a total of 49,870 users with
29,464 males, 16,597 females, and 3,809 gender-unknown
3.5. Novel methodological features of our study users). To the best of our knowledge, this is by far proba-
With respect to methodological issues, one of the great- bly the largest sample ever used in a study of musical prefer-
est challenges in the study of musical preferences are the ences, and certainly the largest one for the population of Chi-
constraints set by genre-related approaches [29]. Here, self- nese music listeners. The statistical power yielded by such
rated musical preferences are given to a list of items charac- a large sample size warrants a reliable detection of curvilin-
terized by a priori specified musical genres. However, as has ear trends that may otherwise remain uncovered in smaller
often been pointed out, genres are only made-up labels that samples (for a related discussion, see Ref. [28]).
have been invented by the music industry to broadly catego-
rize music [29] and as such are rather ill-defined for describ- 4. General discussion and future research directions
ing individual musical preferences [27, 81]. In our present
study, we have tried to circumvent this problem by using The advent of online music-based social networks and the
the genre information that is actually constructed by collab- increasing accessibility of data on virtual behaviors have en-
orative tagging of NCM users. Instead of using predefined abled unprecedented large-scale studies of human cultural
genres for each considered track in our dataset, the genre dynamics. Our present study investigated the formation of
17
 CI CA CU
(a) <D >, Province Level (b) D , Province Level (c) D , Province Level

Community Diversity

r = 0.35 r = 0.60 r = 0.48

P = 0.102 P = 0.0027 P = 0.021

-8 -7 -7
Slope = 9.6×10 Slope = 3.3×10 Slope = 2.3×10

CI CA CU
(d) <D >, City Level (e) D , City Level (f) D , City Level

r = 0.12 r = 0.22 r = 0.10

P = 0.056 P = 0.00035 P = 0.102

per Capita Disposable Income (in CNY)

Figure 14: (Color online) The relationship between the average individual community diversity hDCI i (panel (a)), the aggregated community diversity DCA
(panel (b)), and the within-group community diversity DCU (panel (c)) and the per capita disposable income at the province level. Panels (d), (e), and (f),
respectively, show the relationships between each diversity measure and the per capita disposable income at the city level. The currency unit of the per capita
disposable income was CNY.

(a) (b)

Language

Theme

*
Genre

**
**
T

Scenario

**
*
Emotion

**
** All

Age 12 ~ 18
Theme
Age 19 ~ 25

Age 26 ~ 40

-1.0 -0.5 0.0 0.5

Correlation per capita Disposable Income (in CNY)

Figure 15: (Color online) (a) The correlations between the gender differences in musical preferences (measured by the tag-based KLD ∆T ) and the disposal
income of each province for different tag classes. Correlations are shown for all users, the users aged between 12 and 18, 19 and 25, and 26-40. (b) Gender
differences (∆T ) between users in each province as a function of the per capita disposable income of each province. The currency unit of the per capita
disposable income was CNY.

18
musical preferences and their variability with demographic music tastes are greatly shaped by the musical environments
factors by analyzing 30,562,590 music playlists, 2,247,960 and events occurring in the teen ages (especially around the
songs, and behaviors of 3,028,351 users of the NCM on- age of 13), and at the time of the early adulthood (ages 20-
line music-based community, one of the largest online music 25), representing the two major transitions in the formation
platforms in China. of musical preferences. Individuals then tend to further ex-
Combining the tools from information sciences, statisti- plore different types of music but also to develop more stable
cal physics, and complex networks theory within the context and personalized styles following the age of 25.
of Big Data analysis, we unveiled several systematic pat- However, with respect to these critical periods and major
terns in the formation and evolution of musical preferences transitions in the development of music preference behav-
in China. We further employed a series of statistical mea- iors, our analyses reveal a remarkable discrepancy between
sures to quantitatively characterize different variables related the conclusions of earlier offline (e.g. [60]) and more re-
to both individual-level and global musical preferences in the cent online studies of musical preferences (e.g. [61]). More
studied population, including users’ tagging diversity, com- specifically, previous offline studies, that were also typically
munity diversity, individual and global musical preferences, limited to small sample sizes, found an inverted U-shaped
the decay pattern of music influence, and user’s sensitivity to pattern characterizing the development of musical prefer-
music. We then modeled the obtained empirical data distri- ences, with a peak observed around the age of 24 [60]. On
butions identifying various features of musical preferences, the other hand, large-scale studies of preference behaviors
including peak sensitivity to popular music that was reached in online music-based communities [33, 61], including our
at the age of 13, thereby strikingly matching recent findings present investigation, suggest that the period of maximum
observed for US American users of Spotify, another online sensitivity to music starts much before the early adulthood,
music-based community [61]. We finally performed a series in the early teens of music listeners. Since these differences
of correlational analyses to address the relationships between may be due to many factors (e.g. differences in age-ranges of
different music preference variables and demographic factors participants, differences in the number of investigated music
such as age, gender, and economic development. tracks and genres etc.), it will remain an evident challenge
The striking similarity with recent online music commu- for future research to identify the underlying origin of this
nity analyses of non-Chinese music listeners (e.g. [61]) al- discrepancy.
lows us to assume some level of cross-cultural universality Critical periods in the development of musical preferences
in the formation of musical preferences, but not in the re- reported in our study are well supported by interactionist the-
lationship between economic development and genre-based ories (e.g. [62]), postulating that musical preferences emerge
musical tastes (cf. [31]). By examining the features of the and develop in accordance with listeners’ basic psychologi-
user-music bipartite network, we observed a stable tempo- cal needs and emotions, their reinforcement, and as the re-
ral trend of musical tastes among Chinese listeners of NCM: flection of the emerging social identities that can be traced
The musical preferences decay slowly in a power-law-like across different developmental phases. For example, the val-
fashion. Such specific patterns have been identified in some ley of the within-group community diversity that was ob-
online social-media behaviors, e.g. in the outbreak of Inter- served in our study for the age period 20-25 corresponds
net memes and in the spreading of popular topics [63, 64]. largely to the more homogeneous and stable phase of the
However, the observed decay of musical preferences in university campus life, whereas the rapidly growing within-
our study was much slower than what has previously been group diversities after the age of 25 are more likely to be
reported for other online behaviors. Namely, the complete associated with the later social influences prevalent in the
decay process in our present study spreads over the extended emerging professional careers.
period of 20 years, suggesting a remarkably stretched phase Our study has also revealed strong gender differences in
of musical taste formation and new music discovery, accom- the development of musical tastes. In general, the within-
panied by a slow taste freezing in the population of Chinese group differences in musical preferences of women are larger
music listeners. In other words, the change of music pref- than those of men, even though their genre-related individual
erences is often a generational phenomenon in China. This diversity is lower than in male users. These findings largely
interesting finding raises the question of whether the char- break through the individual perspectives reported in previ-
acteristics of musical preferences are generalizable to other ous studies [38], but they also expand our understanding of
types of nonmusical preference behaviors, such as those for musical taste formation towards the multidimensional, socio-
films, dance, fashion, etc. Further comparative studies across psychological-technological perspective, revealing an intrin-
different cultural products will be necessary to address this sic complexity of gender trends in online music preference
more global level of potential universality of cultural prefer- behaviors.
ences. Importantly, we observed the strongest gender differences
We further mapped the age trends of each tag and classi- among those online music listeners who were in their teens
fied them into nine basic age modes. In terms of individual (see Fig. 11). This is in contrast with a previous study of
musical sensitivity, we found that teenagers are more sensi- US users of Spotify (e.g. [77]), in which male and female
tive to music than adults, which is consistent with recent Big teenagers were found to share highly similar listening be-
Data analyses (e.g. [61]). After the age of 25, the individ- haviors. This discrepancy between adolescent Chinese and
ual preference diversities for both tags and communities start US music listeners suggests that unlike age, the role that
to decline, and the within-group diversities rapidly increase gender plays in music preference formation is rather culture-
for most of the investigated cases. In other words, individual specific. However, since the explorative study of Kalia [77]
19
did not consider a detailed structure of musical preferences and evaluate their potential usage in online music recommen-
and largely focused on music popularity, further compara- dation systems.
tive and more comprehensive analyses will be necessary to Datasets collected in our study could also be employed
draw conclusive statements about cross-cultural gender dif- to calibrate future large-scale agent-based models [83, 84],
ferences in the formation of musical preferences. that would then simulate various aspects of users’ preference
Our study also identified a relationship between musical behaviors which are otherwise hidden or difficult to extract
preferences and regional economic development. We found from music-based social networks. For example, relative
that economic development is mainly related to within-group contributions of different intervening variables such as fac-
diversity in music preferences, but not to diversity at an tors related to online technology use or social influence ex-
individual level, suggesting thus the growth of some mi- erted among users could be estimated by using agent-based
nority music circles in China. The negative correlations simulations. Such models could further investigate how at-
between economic indicators and tag strengths for melan- tention diffusion [76], opinion flow [85, 86], or other spread-
cholic/negative emotions suggest a possible positive influ- ing processes [87] in online social networks shape the for-
ence of economic development on listeners’ well-being and mation of both individual and global musical preferences.
life outlook. Moreover, the generally stronger correlations in Due to their generative character, agent-based models would
female users indicate that the musical preferences of women enable us to understand not only what types of global pat-
could be more sensitive to economic development. terns can possibly exist in online music-based communities,
Interestingly enough, the observed negative correlation but more importantly, they could potentially inform us about
between gender differences in music preferences and eco- how these patterns actually emerge and evolve from multiple
nomic indicators in our study (Fig. 15(b)) is in contrast with local preference behaviors and interactions occurring among
a recently observed positive relationship between gender dif- millions of connected individuals.
ferences in preference behaviors and economic development In the meantime, while such ambitious large-scale agent-
[82], suggesting that these gender differences can actually based modeling is still in progress, next studies should go
widen with a more egalitarian availability of resources and beyond mere correlational analyses and address potential in-
the associated enhanced formation of gender-specific prefer- teraction effects between age, gender, and economic devel-
ences. However, we note here that this positive relationship opment, e.g. via analyses within the context of the structural
between gender differences in preferences and economic fac- equation modeling framework (see e.g. [28]) or linear mixed
tors [82] was not observed for musical preferences but only effects models [88].
for behaviors that involved economic decision making pro- Finally, it could also be highly useful for the future playlist
cesses, e.g. for preferences in risk-taking or in various types generation and recommendation system design to investigate
of costly cooperative behaviors. It is therefore possible, as if any changes in the patterns of musical preferences can be
our current study seems to suggest, that preference behav- observed in the ongoing post-COVID-19 era relative to the
iors which do not involve economic decisions (such as pref- previously identified, pre-COVID patterns of musical tastes
erences for different musical genres) tend to show a negative that have been reported in our present paper. Indeed, in the
correlation with economic development, because gender dif- wake of the ongoing pandemic, music has served the role of
ferences in non-costly behaviors are more likely to diminish a connecting tissue [89] bringing together millions of people
in the presence of a gender-equal access to resources. The in times when direct physical contact was rare and explic-
observed decline in gender differences in music preferences itly discouraged as an integral part of epidemic mitigation
with an increasing economic development in our present pa- strategies. It would therefore be promising to investigate if
per is supported by the social role theory [65], according preference behaviors and their relationship to demographic
to which gender differences in preference behaviors should indicators might have changed since the outbreak, and more-
weaken in more developed regions, where higher economic over, which dynamic trends can be observed over the whole
development and the associated declining relevance of con- time course of the pandemic.
ventional gender roles are expected to narrow gender gaps in
preference behaviors.
In spite of its largely explorative nature and a variety of 5. Conclusions
methodological challenges, Big Data technology has a wide
ranging applicability in a number of domains [51], where This study presents a Big Data analysis of global patterns
it can successfully complement various predictive modeling in musical taste formation in China based on real behavioral
techniques. One potential area of application of our current data obtained from millions of social media users. Employ-
study is music recommendation [46] in online music stream- ing advanced methods from information sciences, statistical
ing services and music-based online communities such as physics, and network theory, we discovered universal age
the NCM or Spotify. For example, similarly to a previous and gender trends in musical preferences formation, as well
proposal [27], the information obtained in our study from as global patterns of musical sensitivity and genre-based di-
user-based genre-tagging of music tracks and from other on- versity in musical tastes.
line behaviors may help in capturing the latent structure of Our analyses revealed two major transitions in the devel-
users’ musical preferences much better than the usually em- opment of musical preferences: One is associated with the
ployed procedures with artificially labeled music genres. Fu- peak sensitivity to music influence (occurring around the age
ture studies should therefore assess the capacity of our anal- of 13), and the other is related to musical taste freezing, i.e.
yses for predicting individual online preference behaviors, the decaying trend in the formation of musical preferences
20
(around the age of 25). Similar to earlier studies with partic- and the studies that will follow, will help us lay the founda-
ipants from US and UK, we observed that whilst the interest tion for a more general, quantitative, and integrative theory
in some musical genres and listening contexts decreased in of musical preferences formation, that will advance our un-
adulthood of Chinese music listeners, their preferences for derstanding of universal properties of human culture and its
other musical dimensions actually increased with age. evolutionary dynamics.
Further comparisons with recent analyses of online pref-
erence behaviors of Western listeners allowed us to surmise Acknowledgments
that these features in the development of music preferences
may be universal, shared across many different cultures and This work was supported by the National Natural Science
times. These findings lend support to interactionist theories Foundation of China (grant no. 62073112), the Centre for
[62] that highlight the importance of individual-environment Communication Research at City University of Hong Kong
relationships and emphasize preferences for those musical (grant no. 9360120), and the Hong Kong Research Grants
surroundings that reinforce the traits of an individual’s per- Council (grant no. 11505119).
sonality, emotions, and values, and reflect the emerging so-
cial identity of an individual. References
In addition to universal properties in musical taste forma-
[1] M. Park, J. Thom, S. Mennicken, H. Cramer, M. Macy, Global music
tion, we identified some culture-specific features including
streaming data reveal diurnal and seasonal patterns of affective prefer-
eight distinct communities of music listeners in the popu- ence, Nature Hum. Behav. 3 (2019) 230.
lation of NCM users, and a relationship between regional [2] P.E. Savage, S. Brown, E. Sakai, T.E. Currie, Statistical universals
economic development and preferences for specific music reveal the structures and functions of human music, Proc. Natl Acad.
Sci. USA 112 (2015) 8987–8992.
genres in China which has not been reported previously for [3] A. Ravignani, T. Delgado, S. Kirby, Musical evolution in the lab ex-
Western music listeners. Specifically, the correlation pat- hibits rhythmic universals, Nature Hum. Behav. 1 (2016) 1–7.
terns typically observed in US music listeners between some [4] D.M. Greenberg, S.J. Wride, D.A. Snowden, D. Spathis, J. Potter, P.J.
musical genres (e.g. rap, soul, or reggae) and economic Rentfrow, Universals and variations in musical preferences: A study
of preferential reactions to Western music in 53 countries, J. Pers. Soc.
indicators were not found in the Chinese NCM population Psychol. 122 (2022) 286.
of listeners, as the values represented by these genres are, [5] D.J. Levitin, The world in six songs: How the musical brain created
in accordance with predictions of the social identity theory human nature, Plume, New York, 2008.
[80], generally less shared among Asian music consumers. [6] P.N. Juslin, From everyday emotions to aesthetic emotions: Towards a
unified theory of musical emotions, Phys. Life Revs. 10 (2013) 235–
We also found that in their adolescence, relatively to their 266.
male counterparts, female users of music-based social media [7] G. Bernatzky, M. Presch, M. Anderson, J. Panksepp, Emotional foun-
in China were characterized by more heterogeneous music dations of music as a non-pharmacological pain management tool in
preference behaviors that were also stronger associated with modern medicine, Neurosci. Biobehav. Revs. 35 (2011) 1989–1999.
[8] E.G. Schellenberg, Music lessons enhance IQ, Psychol. Sci. 15 (2004)
economic factors. This gender gap in the adolescence was 511–515.
observed to decrease with the onset of the early adulthood [9] E.W. Large, On synchronizing movements to music, Hum. Movement
and with the increasing economic development. Sci. 19 (2000) 527–566.
[10] P. Ball, Facing the music, Nature 453 (2008) 160–162.
Our findings thus not only provide the large-scale empiri- [11] A.E. Coca, L. Zhao, Musical rhythmic pattern extraction using rele-
cal evidence for the existence of universal patterns in musical vance of communities in networks, Inf. Sci. 329 (2016) 819–848.
taste development, but they also reveal a rich and novel set of [12] E. Nakamura, K. Kaneko, Statistical evolutionary laws in music styles,
features characterizing the emergence of culture-dependent Sci. Rep. 9 (2019) 15993.
[13] F. Pachet, P. Roy, Markov constraints: Steerable generation of Markov
preference behaviors, that can provide further insights into sequences, Constraints 16 (2011) 148–172.
regional cultural diffusion and music popularization pro- [14] T.J. Tighe, W.J. Dowling, Psychology and music: The understanding
cesses. Moreover, our findings reveal remarkable differences of melody and rhythm, Psychology Press, New York, 2014.
between earlier offline and more recent online studies of mu- [15] A.S. Cowen, X. Fang, D. Sauter, D. Keltner, What music makes us
feel: At least 13 dimensions organize subjective experiences associ-
sical preferences, that may be due to differences in the stud- ated with music across different cultures, Proc. Natl Acad. Sci. USA
ied sample sizes but also due to the ever-growing availability 117 (2020) 1924–1934.
of music and the increasing sophistication of its use in online [16] D.J. Levitin, P. Chordia, V. Menon, Musical rhythm spectra from Bach
to Joplin obey a 1/f power law. Proc. Natl Acad. Sci. USA 109 (2012)
networking technologies. Importantly, our study has shown
3716–3720.
that applying the network theoretic formalism to large-scale [17] J. Su, P. Zhou, Use of Gaussian process to model, predict and explain
behavioral datasets can serve as a powerful tool for analyzing human emotional response to Chinese traditional music, Adv. Com-
the community structure and the diversity of cultural links in plex Syst. 24 (2021) 2250001.
[18] P. Ball, The music instinct: How music works and whywe can’t do
online music platforms such as the NCM, enabling us to un- without it, Bodley Head, London, UK, 2010.
cover musical preference trends and their global features that [19] N. Askin, M. Mauskapf, What makes popular culture popular? Prod-
otherwise would have remained undetected. uct features and optimal differentiation in music, Am. Sociol. Rev. 82
These findings contribute to our multidisciplinary under- (2017) 910–944.
[20] J. Serrá, Á. Corral, M. Boguñá, M. Haro, J.L. Arcos, Measuring the
standing of musical preference behaviors in online social evolution of contemporary western popular music, Sci. Rep. 2 (2012)
networks, with implications for a wide variety of applied do- 1–6.
mains such as assessment of online content popularity, user [21] L. Trainor, The emotional origins of music, Phys. Life Revs. 7 (2010)
44.
profiling, community detection in online music platforms,
[22] B. Merker, I. Morley, W. Zuidema, Five fundamental constraints on
and the design of recommendation systems for online so- theories of the origins of music, Phil. Trans. R. Soc. B 370 (2015)
cial media. Taken together, we hope that our present results, 20140095.

21
[23] W.T. Fitch, The biology and evolution of music: A comparative per- complex networks, J. Syst. Sci. Complex. 27 (2014) 650–665
spective, Cognition 100 (2006) 173–215. [50] D. Garcı́a, D. Tanase, Measuring cultural dynamics through the Euro-
[24] H. Miton, T. Wolf, C. Vesper, G. Knoblich, D. Sperber, Motor con- vision song contest, Adv. Complex Syst. 16 (2013) 1350037.
straints influence cultural evolution of rhythm, Proc. R. Soc. B 287 [51] H. Wang, Z. Xu, H. Fujita, S. Liu, Towards felicitous decision making:
(2020) 20202001. An overview on challenges and trends of Big Data, Inf. Sci. 367-368
[25] P.E. Savage, P. Loui, B. Tarr, A. Schachner, L. Glowacki, S. Mithen, (2016) 747–765.
W.T. Fitch, Music as a coevolved system for social bonding, Behav. [52] Z.-Q. You, X.-P. Han, T. Hadzibeganovic, The role of research effi-
Brain Sci. 44 (2021) e59. ciency in the evolution of scientific productivity and impact: An agent-
[26] A. Lamont, D. Hargreaves, Musical Preferences, in Routledge Inter- based model, Phys. Lett. A 380 (2016) 828–836.
national Handbook of Music Psychology in Education and the Com- [53] D.N. Rockmore, C. Fang, N.J. Foti, T. Ginsburg, D.C. Krakauer, The
munity, Routledge, pp. 131–145, 2021. cultural evolution of national constitutions, J. Assoc. Inf. Sci. Technol.
[27] P.J. Rentfrow, L.R. Goldberg, D.J. Levitin, The structure of musical 69 (2018) 483–494. https://doi.org/10.1002/asi.23971
preferences: A five-factor model, J. Pers. Soc. Psychol. 100(6) (2011) [54] M. Scheffer, I. van de Leemput, E. Weinans, J. Bollen, The rise and
1139–1157. fall of rationality in language, Proc. Natl. Acad. Sci. USA 118 (2021)
[28] A. Bonneville-Roussy, P.J. Rentfrow, M.K. Xu, J. Potter, Music e2107848118.
through the ages: Trends in musical engagement and preferences from [55] S. DeDeo, Using big data to track major shifts in human cognition,
adolescence through middle adulthood, J. Pers. Soc. Psychol. 105 Proc. Natl. Acad. Sci. USA 119 (2022) e2121300119.
(2013) 703–717. [56] K.P. Subbu, A.V. Vasilakos, Big Data for context aware computing –
[29] D.M. Greenberg, S. Baron-Cohen, D.J. Stillwell, M. Kosinski, P.J. Perspectives and challenges, Big Data Research 10 (2017) 33–43.
Rentfrow, Musical preferences are linked to cognitive styles, PLoS [57] T.-Q. Peng, J.J.H. Zhu, Sophistication of Internet usage (SIU) and its
ONE 10 (2015) e0131151. attitudinal antecedents: An empirical study in Hong Kong, Comput.
[30] H. Li, X.-P. Han, L. Lü, Z. Pan, Measuring diversity of music tastes in Hum. Behav. 27 (2011) 421–431.
online musical society, Int. J. Mod. Phys. C 29 (2018) 1840006. [58] D.J. Zhang, M. Hu, X. Liu, Y. Wu, Y. Li, NetEase cloud music data,
[31] C. Mellander, R. Florida, P.J. Rentfrow, J. Potter, The geography of Manuf. Serv. Oper. Manag. 24 (2022) 275–284.
music preferences, J. Cult. Econ. 42 (2018) 593–618. [59] M.A. Keijzer, M. Mäs, A. Flache, Communication in online so-
[32] M. Park, I. Weber, M. Naaman, S. Vieweg, Understanding musical cial networks fosters cultural isolation, Complexity (2018) 9502872.
diversity via online social media, Proc. Intl. AAAI Conf. Web Soc. https://doi.org/10.1155/2018/9502872
Media, AAAI Press, pp. 308–317, 2016. [60] M.B. Holbrook, R.M. Schindler, Some exploratory findings on the
[33] S.F. Way, S. Gil, I. Anderson, A. Clauset, Environmental changes and development of musical tastes, J. Consum. Res. 16(1) (1989) 119–
the dynamics of musical identity, Proc. Intl. AAAI Conf. Web Soc. 124.
Media, AAAI Press, pp. 527–536, 2019. [61] S. Stephens-Davidowitz, The songs that bind, New York Times,
[34] L. Xu, Y. Zheng, D. Xu, L. Xu, Predicting the preference for sad mu- February 11, SR9, 2018.
sic: The role of gender, personality, and audio features, IEEE Access [62] W.B. Jr. Swann, P.J. Rentfrow, J.S. Guinn, Self-verification: The
9 (2021) 92952–92963. search for coherence, In M. Leary, J. Tagney (Eds.), Handbook of self
[35] K.R. Fricke, D.M. Greenberg, P.J. Rentfrow, P.Y. Herzberg, Measur- and identity, Guilford Press, New York, 2002.
ing musical preferences from listening behavior: Data from one mil- [63] R. Crane, D. Sornette, Robust dynamic classes revealed by measuring
lion people and 200,000 songs, Psychol. Music 49 (2021) 371–381. the response function of a social system, Proc. Natl Acad. Sci. USA
[36] D. Rawlings, V. Ciancarelli, Music preference and the five-factor 105(41) (2008) 15649–15653.
model of the NEO personality inventory, Psychol. Music 25(2) (1997) [64] J. Leskovec, L. Backstrom, J. Kleinberg, Meme-tracking and the dy-
120–132. namics of the news cycle, In Proc. 15th ACM SIGKDD Intl Conf.
[37] A. Langmeyer, A. Guglhor-Rudan, C. Tarnai, What do music pref- Knowl. Discovery and Data Mining (KDD ’09), pp. 497–506, ACM,
erences reveal about personality? A cross-cultural replication using New York, 2009.
self-ratings and ratings of music samples, J. Individual Differences 33 [65] A.H. Eagly, W. Wood, The origins of sex differences in human behav-
(2012) 119–130. ior: Evolved dispositions versus social roles, Am. Psychol. 54 (1999)
[38] A. Bonneville-Roussy, D. Stillwell, N. Kosinski, J. Rust, Age trends in 408–423.
musical preferences in adulthood: 1. Conceptualization and empirical [66] Z. Zhu, J. Su, L. Kong, Measuring influence in online social network
investigation, Musicae Scientiae 21(4) (2017) 369–389. based on the user-content bipartite graph, Comput. Hum. Behav. 52
[39] A. Bonneville-Roussy, J. Rust, Age trends in musical preferences in (2015) 184–189.
adulthood: 2. Sources of social influences as determinants of prefer- [67] M.E.J. Newman, Detecting community structure in networks, Eur.
ences, Musicae Scientiae 22 (2017) 175–195. Phys. J. B 38(2) (2004) 321–330.
[40] A. Bonneville-Roussy, E.T. Tuomas, Age trends in musical prefer- [68] K. Berahmand, A. Bouyer, A link-based similarity for improving com-
ences in adulthood: 3. Perceived musical attributes as intrinsic deter- munity detection based on label propagation algorithm, J. Syst. Sci.
minants of preferences, Musicae Scientiae 22 (2017) 394–414. Complexity 32 (2019) 737–758.
[41] S. Shin, J. Park, On-chart success dynamics of popular songs, Adv. [69] Z. Xu, X. Rui, J. He, Z.X. Wang, T. Hadzibeganovic, Superspreaders
Complex Syst. 21 (2018) 1850008. and superblockers based community evolution tracking in dynamic
[42] D. Guerrero, P. Rivera, G. Febres, C. Gershenson, Towards a mea- social networks, Knowl.-Based Syst. 192 (2020) 105377.
sure for characterizing the informational content of audio signals and [70] R. Kumar, R. Dohare, Formalising and detecting community struc-
the relation between complexity and auditory encoding, Entropy 23 tures in real world complex networks, J. Syst. Sci. Complexity 34
(2021) 1613. (2021) 180—205.
[43] Q. Guo, L. Ji, J.-G. Liu, J. Han, Evolution properties of online user [71] Z. Li, R.-S. Wang, S. Zhang, X.-S. Zhang, Quantitative function and
preference diversity, Physica A 468 (2017) 698–713. algorithm for community detection in bipartite networks, Inf. Sci. 367-
[44] Y. Liu, J. Wang, Y. Jiang, J. Sun, J. Shang, Identifying impact of in- 368 (2016) 874–889.
trinsic factors on topic preferences in online social media: A nonpara- [72] V.D. Blondel, J.-L. Guillaume, R. Lambiotte, L.E. Etienne, Fast un-
metric hierarchical Bayesian approach, Inf. Sci. 423 (2018) 219–234. folding of communities in large networks, J. Stat. Mech. 2008 (10)
[45] E. Arrigo, C. Liberati, P. Mariani, Social media data and users’ prefer- (2008) P10008.
ences: A statistical analysis to support marketing communication, Big [73] R. Lambiotte, P. Panzarasa, Communities, knowledge creation, and
Data Research 24 (2021) 100189. information diffusion, J. Informetr. 3(3) (2009) 180–190.
[46] S.K. Lee, Y.H. Cho, S.H. Kim, Collaborative filtering with ordinal [74] N. Eagle, M.W. Macy, R. Claxton, Network diversity and economic
scale-based implicit ratings for mobile music recommendations, Inf. development, Science 328 (2010) 1029–1031.
Sci. 180 (2010) 2142–2155. [75] S. Kullback, R.A. Leibler, On information and sufficiency, Ann. Math.
[47] J.-H. Wang, Q. Guo, K. Yang, Y.-L. Zhang, J. Han, J.-G. Liu, Popu- Stat. 22(1) (1951) 79–86.
larity and user diversity of online objects, Physica A 461 (2016) 480– [76] C. Candia, C. Jara-Figueroa, C. Rodriguez-Sickert, A.-L. Barabási,
486. C.A. Hidalgo, The universal decay of collective memory and attention,
[48] K. Gu, Y. Fan, Z. Di, Signed PageRank on online rating systems, J. Nature Hum. Behav. 3 (2019) 82–91.
Syst. Sci. Complex. 35 (2022) 58–80 [77] A. Kalia, Music was better back then: When do we stop keeping up
[49] P. Wang, C. Tian, J. Lu, Identifying influential spreaders in artificial with popular music?, Skynet & Ebert Blog, April 22, 2015.

22
 https://skynetandebert.com/2015/04/22/music-was-better-back-then-
when-do-we-stop-keeping-up-with-popular-music/
[78] R. Hogan, B.W. Roberts, A socioanalytic model of maturity, J. Career
Assess. 12 (2004) 207–217.
[79] K. Zagorski, J. Kelley, M.D. Evans, Economic development and hap-
piness: Evidence from 32 nations, Polish Sociol. Rev. 169 (2010) 3–
19.
[80] H.G. Tekman, N. Hortaçsu, Music and social identity: Stylistic iden-
tification as a response to musical style, Int. J. Psychol. 37 (2002)
277–285.
[81] R. Lambiotte, M. Ausloos, On the genre-fication of music: A percola-
tion approach, Eur. Phys. J. B 50 (2006) 183–188.
[82] A. Falk, J. Hermle, Relationship of gender differences in preferences
to economic development and gender equality, Science 362 (2018)
eaas9899.
[83] E. Bonabeau, Agent-based modeling: Methods and techniques for
simulating human systems, Proc. Natl Acad. Sci. USA 99 (2002)
7280–7287.
[84] T. Hadzibeganovic, P.B. Cui, Z.X. Wu, Nonconformity of cooperators
promotes the emergence of pure altruism in tag-based multi-agent net-
worked systems, Knowl.-Based Syst. 171 (2019) 1–24.
[85] K. Lichtenegger, T. Hadzibeganovic, The interplay of self-reflection,
social interaction and random events in the dynamics of opinion flow
in two-party democracies, Int. J. Mod. Phys. C 27(5) (2016) 1650065.
[86] W. Su, X. Wang, G. Chen, Y. Yu, T. Hadzibeganovic, Noise-based
synchronization of bounded confidence opinion dynamics in hetero-
geneous time-varying communication networks, Inf. Sci. 528 (2020)
219–230.
[87] Q.C. Wu, T. Hadzibeganovic, An individual-based modeling frame-
work for infectious disease spreading in clustered complex networks,
Appl. Math. Model. 83 (2020) 1–12.
[88] D.M. Bates, M., Mächler, B.M. Bolker, S.C. Walker, Fitting linear
miexed-effectsmodels using lme4, J. Stat. Software 67 (2015) 1–48.
[89] D.M. Greenberg, J. Decety, I. Gordon, The social neuroscience of mu-
sic: Understanding the social brain through human song, Am. Psy-
chol. 76 (2021) 1172–1186.

23