Emerging research suggests that cultural richness and complexity intensify with population size. Yet the mechanism underlying this phenomenon remains unclear: Do populated areas exhibit more cultural diversity simply due to there being a larger spectrum of individuals with varied backgrounds, or does the urban environment itself stimulate individuals to explore a wider variety of cultural experiences, raising the population's baseline? To decipher this, we leverage a large-scale dataset of 69 million music listening events from the real world, examining the listening patterns of over 408 thousand unique individuals across 96 regions in France. Our study presents a dual perspective on diversity by (1) measuring one's diversity of musical consumption by evaluating the breadth of their music listening history, and (2) assessing the shared repertoire among individuals as a collective. We found that both individual and collective levels of musical consumption diversity increase with population size. This trend held true when segmenting the population by gender and age groups, while a gender-speci昀椀c divergence in consumption appeared from a particular age. We further delineate potential confounding variables to consider in future research aimed at identifying causal pathways, presenting this model using a Directed Acyclic Graph (DAG). Together, our preliminary work represents a crucial step towards unravelling the complexity of cultural diversity and its ties to population dynamics.
With our world becoming increasingly interconnected, the incredible variety of human cultures emerges as a prominent aspect of our collective existence. Throughout history, from primal cave paintings to the triumph of modern architecture, every artefact, technology, and tradition paints a unique hue onto our collective cultural spectrum.
Contemporary research has begun to apply computational techniques to understand the
constant 昀氀ux of cultures, translating it into discernible patterns and delving into the intricate
processes that shape our shared identity. A plethora of methods have been utilized in this
endeavour, including computer simulations 6[, 7], experimental studies [
However, empirically testing these relationships remains a signi昀椀cant challenge since
understanding the complex nature of cultural dynamics requires data that is high in temporal
and spatial resolutions, capturing both micro and macro levels. Computational models provide
valuable theoretical underpinning but o昀琀en portray speculative scenarios that necessitate
realworld data for validation 9[]. Our study addresses this gap by assembling and analyzing an
extraordinarily large dataset of music listening events, featuring 69 million streams from over
408 thousand unique individuals living in France. This 昀椀nely-grained behavioural data, both
temporally and spatially, provides a natural experiment to investigate the in昀氀uence of
population size and demographic composition on cultural diversity at both individual and collective
scales. Music serves as an ideal test-bed for examining such phenomenon, with its universal
prevalence [
Just as an ecologist might analyze the biodiversity of species within a particular habitat,
we study individuals’ music-listening behaviour within speci昀椀c geographical regions to
understand the diversity of musical consumption. According to prior research that observed elevated
levels of diversity with population size in various domains, including language28[] and
music [
In one hypothetical scenario, increased cultural diversity in densely populated areas, like capital cities, might stem from the convergence of diverse individuals that form a cultural melting pot. In assessing the diversity of these regions, we may register increased diversity simply due to the higher likelihood of encountering individuals from varied backgrounds. Here, the observed regional diversity would then largely be attributed to demographic diversity, meanwhile the extent of an individual’s breadth of consumption may be relatively isolated from regional in昀氀uences; instead, constrained by the general capacity of musical variety one might engage with.
Alternatively, the catalysts for diversity might lie within principles of consumer behaviour and exposure. Inhabitants of populous regions, continuously surrounded by a variety of cultural o昀erings and privileged with abundant access to cultural events, might gradually incorporate this variety into their personal tastes. This continuous exposure to, and easy access to diverse cultural manifestations could subtly mould individual preferences, encouraging exploration and adoption of a broader range of cultural experiences. Over time, these in昀氀uences might alter the population’s baseline, leading to a higher level of diversity in dense metropolitan areas.
To this end, our study seeks to untangle the potential pathways towards cultural diversity, considering diversity at two levels: (1) an individual’s diversity of musical consumption based on their listening pro昀椀le, and (2) a collective’s diversity based on the alignment of consumption between individuals. We present preliminary 昀椀ndings on how demographic factors and population size play a role on these two accounts. Importantly, we outline our future plans to go beyond mere associations and comprehend the causal mechanisms shaping cultural diversity. For this purpose, we have collected data on potential confounding variables and represent the model we aim to test with a Directed Acyclic Graph (DAG). This paper thus sets the stage for a deeper understanding of the interplay between demographics and consumption behaviour, and their in昀氀uence on cultural diversity. To bene昀椀t the community and future research, we plan to openly make available the aggregated data at the regional level and socio-demographic data gathered from various sources, along with all the statistics computed for the analysis.
Our research was conducted on Deezer, a globally available music streaming platform present in 187 countries, boasting a catalogue of over 90 million tracks. The user’s listening behaviour is captured with comprehensive logs, including the date and time of song playback, duration, the listener’s self-reported age and gender, preferred language, type of streaming device and internet connection, and geographical coordinates derived from the IP address.
Initially, we gathered all music streaming events that happened in France between April 3rd and 30th, 2023, a four-week period ensuring a balanced representation of weekdays. To reduce noise and potential biases in our data, we implemented various stringent 昀椀lters. First, we limited our study to tracks played longer than 30 seconds. Second, we considered streams solely over local area networks (LAN) for robust geolocation. Finally, since our study relies on accurate user-to-region attribution, we developed an additional 昀椀lter to exclude users with large travel radius. To do this, we 昀椀rst derived each user’s listening locations using latitude and longitude data extracted from their IP addresses, excluding users with more than 昀椀ve distinct IP addresses. We then determined each user’s travel radius over the four-week study period by calculating the haversine distance between all pairs of their geographical coordinates. We applied a cut-o昀 at a median distance of 50km (see Figure 1B), which was a reasonable threshold to strike a balance between geographical speci昀椀city and retaining a substantial portion of the user sample. This process excluded 32% of the users and we translated the remaining users’ geographical coordinates into NUTS3 leve1l regions using the ‘gioscoR’ R package, which formed the basis of our regional groupings. To test if these geographical 昀椀ltering might bias the result (e.g. users who travel far to work from home or students who live away from home), we made another sample of users without the geolocation 昀椀lters, which replicated the patterns reported 1Nomenclature of Territorial Units for Statistics or NUTS is a geocode standard for referencing the administrative divisions of countries for statistical purposesh.ttps://ec.europa.eu/eurostat/web/nuts here.
Our 昀椀nal dataset encompassed over 69 million unique listening events from 408 thousand individual users (see Figure1A for all users projected on the map). This data was augmented with statistics on regional population size, median age, and gender ratio gathered from Eurostat.2 When compared across the regional population of France, our dataset’s demographics exhibited strong correlations in terms of regional population size ℎ( = 0.97, < .001) and median age ( ℎ = 0.60, < .001). However, when comparing to the entire French population as a whole, there were 25% more young users on Deezer (between ages 15-30) while 30% less among older (between ages 51 to 80), with the mid-age group being similar in their proportions (3% more on Deezer user base). Regarding self-reported gender, the Deezer users were skewed towards the male demographic (34% more males) compared to France’s general population (6% more females), with no signi昀椀cant correlation found when comparing gender ratios across regions (ℎ = 0.06, = 0.53). This suggests that, while our dataset e昀ectively represents distribution across France’s region in terms of user numbers and age, platform-based age and gender biases exist.
In this study, we draw two perspectives on musical consumption diversity, at the individual
(within-user) and collective (between-users) levels (see Figure1C). Given that sample size can
bias the estimates of diversity [
To assess an individual’s diversity of musical consumption, we employed the
GeneralistSpecialist score (GS-score), a previously validated metric in user music exploration and
discovery studies [
1 ∑ ∑
⃖⃗ ⋅ ⃗ ‖⃖⃗‖ ⋅ ‖ ⃗ ‖
The GS-score e昀ectively captures the dispersion of a user’s listening behaviour in the general song embedding space, whereby aspecialist would have a more focused preference (e.g. only listens to Jazz), whereas a generalist would exhibit a broader range of listening habits (e.g. listens to Jazz, Classical, and Metal). To make this score consistent in the direction of our collective diversity measure, we inverted the score (1 - GS ( )) and normalized it to a range between 0 to 100, where 100 represents maximal diversity.
To quantify the diversity found for a given population, existing research has commonly applied measures like the Gini coe昀케cient, Simpson’s index, or Shannon’s entropy. However, these measures provide arbitrary scales, making comparisons challenging. As a solution to this, Hill’s number (also commonly known as the e昀ective number of species) has become a method increasingly popular to quantify the diversity of an assemblage, which allows for standardized comparisons 3[, 7]. By treating each song as a ‘specie’ in a given population, we calculate collective diversity of consumption ase昀ective number of songs ( ), expressed as: (1) (2) = (∑ )
=1 1/(1−)
Here, de昀椀nes the order of the Hill’s number, where higher values of emphasize the contribution of rare songs, while lower values of focus on the abundance of popular songs. In our analysis, we set the order of to be 1, which essentially becomes the exponential of the familiar Shannon index. The represents the total number of unique songs, and signi昀椀es the relative abundance of each song.
We begin by examining the dynamics of an individual’s musical consumption, and how these may be in昀氀uenced by age, gender and geographical environment they live in.
We 昀椀rst assessed the evolution of one’s musical consumption diversity with respect to
selfreported age and gender. Our 昀椀ndings were consistent with established literature showing
that one’s musical consumption consolidates mostly during adolescence and less exploration
happens with ageing [
Our next analysis focused on the impact of population size, established by the number of users in each NUTS3 region. We observed a positive and substantial correlation ℎ( = 0.59, = 0.49, < .001) on individual’s consumption diversity as a function of population size, indicating that individuals residing in densely populated regions tend to exhibit a more diverse consumption behaviour. To examine whether the e昀ect is uniformly present across di昀erent gender and age groups, we subdivided the population intoyounger (age 18-30) and older (30+), further strati昀椀ed by gender. Across all subgroups, we found an overall increasing trend in diversity with population size when 昀椀tting the data with the generalized additive model (GAM) (see Figure 2B). GAM is a statistical model that extends the concept of the generalized linear model by allowing for non-linear relationships between the dependent variable and the independent variables. GAMs can capture complex relationships by using smooth functions, such as splines, to model the non-linear components. Together, these results indicate that both demography and geographical environment play crucial roles in shaping one’s breadth of musical consumption.
We now broaden our analysis to the diversity of musical consumption across individuals, effectively investigating how similar or distinct one’s musical listening event might be from one another within the same group (i.e. collective). We have previously assessed an individual’s consumption diversity trajectory as a function of age and gender. Here, we group individuals of the same gender and age, analyzing their co-consumption of music (i.e. the shared repertoire and frequency of songs within the group). High co-consumption indicates shared consumption, whereas low overlap points to unique consumption across individuals. The skewness or uniformity of this distribution serves as our diversity metric, determined using the e昀ective number of songs (see Equation2).
Age and gender e昀ects on collective consumption diversity re昀氀ected similar trends to those observed in individual diversity (see Figure3A). Much like individual diversity, between-user diversity increased with a notable surge at age 23 and peaking around the mid-30s (in other words, when gathering a group of mid 30 years old together in one room, they will have the least amount of shared repertoire of songs). We once again identi昀椀ed gender-based di昀erences, with female listeners displaying more shared consumption within their age group than their male counterparts. Additionally, collective consumption diversity declined with age among female listeners, while it plateaued among males from the mid-30s onwards.
In alignment with the higher individual diversity we observed in more populated regions, we also observed an increasing trend for collective diversity with population siz eℎ( = 0.96, = 0.52, < .001), demonstrating a remarkably clear concave downwards curve across all four subgroups (see Figure 3B). This heightened diversity in populous regions becomes distinctly evident when examining the distribution of users in the user embedding, where users with similar listening habits are closer together in this embedding space. As depicted in Figur3eC, users from the capital area Paris are scattered throughout various parts of the space, implying a higher individuality in their music consumption. In contrast, music listeners from a more rural region such as Hérault cluster more closely together, suggesting a more shared musical listening habit.
Our analysis reveals an elevated level of diversity in musical consumption in densely populated
areas. This enhanced diversity is apparent not only at the individual level, but also within the
broader spectrum of the region’s population. Despite these 昀椀ndings, unravelling the
underlying mechanisms that foster these varied consumption patterns and understanding their
complex intertwined relationships is challenging. The strong link we observe between
consumption diversity with increasing population size could be in昀氀uenced by various factors, ranging
from the con昀氀uence of diverse ethnic groups, and socioeconomic inequalities in accordance
with the ‘cultural omnivore theory’ [
Determining causal relationships among these potential factors would ideally necessitate carefully controlled lab experiments. However, in the context of our study and many others examining social phenomena, it is o昀琀en not feasible nor ethical to design experiments manipulating individuals’ cultural environments. An alternative approach is to rely on observational data while thoroughly considering potential variables that could be confounding and carefully controlling for them with the available modern tools2[9].
To this end, we have collected data to account for various potential factors mentioned in
existing literature. These data were extracted and augmented from various external sources: (i)
self-reported age and gender, (ii) population demographics (population size, immigration,
education, income) obtained from Eurostat and the National Institute of Statistics and Economic
Studies (INSEE) of France, and (iii) information on musical venues gathered using SongKick’s
API3 to estimate the number of active musical events. We utilized these variables to construct
a model for future testing, as shown in Figure4. This model is represented by Directed Acyclic
Graphs (DAGs), which o昀er a clear and e昀케cient approach to identifying, presenting, and
hypothesising the causal relationships between variables [
By studying the behaviour of hundreds of thousands of music listeners in the real world, our study delves into the diversity of cultural consumption drawing two levels of perspectives, examining not only the breadth of an individual’s consumption but also the alignment of consumption as a collective, strati昀椀ed by age, gender, and geographical regions.
Our analysis reveals that both individual and collective levels of musical consumption diversity positively correlate with population size. This suggests that individuals residing in more densely populated regions not only have a broader spectrum of musical consumption but also exhibit more unique consumption pro昀椀les to one another within the population. This trend persists across di昀erent demographics, holding true when the population is segmented by gender and age groups.
It is important to note, however, that our current 昀椀ndings only suggest correlations and do not establish causality. As such, in order to discern causal pathways that cultivate diversity at various levels, we outline potential confounding variables to consider for future research. These variables are incorporated into a model represented as a DAG. In subsequent research, we aim to test this model to determine the direct e昀ects of various factors contributing to the observed diversity.
Our results align with prior research studying the trajectories of musical exploration across an individual’s lifespan, revealing a general consolidation of consumption during adolescence. Yet, a notable observation from our study is the divergence in musical consumption between self-reported male and female listeners from a particular age onward. This disparity could potentially be a cultural phenomenon, or it may stem from other underlying factors that are hidden from our current observation.
In the broader scheme of verifying and generalizing the 昀椀ndings reported in this paper, we intend to replicate our analysis in other countries — Brazil and Germany. Additionally, we aim to openly share the demographic data and indicators we have compiled, along with the aggregated music stream data and corresponding analysis code for reproducibility and transparency.