Depict or Discern? Fingerprinting Musical Taste from Explicit Preferences

2.1 Music preference data collection

In sociology and psychology, collecting declarative data about musical preferences and consumption habits through surveys and interviews is common. Interacting directly with the respondents is advantageous for several reasons. The use of a Likert scale for instance allows to have a deeper understanding of how much respondents do or do not like certain music (Peterson, 1992; Bryson, 1996). Information about context of music consumption can be collected (DeNora, 2000), as well as sociodemographics, that can then be crossed with declared music preferences (Bourdieu, 1984; Peterson, 1992; Bryson, 1996; Coulangeon, 2017; Lahire, 2008). However, the sample of surveyed individuals is usually limited, and the results can be biased as such surveys are often run either in a specific country, or on a specific social group, like students for example (Delsing et al., 2008; Brown, 2012; Langmeyer et al., 2012). Additionally, the respondents may find it difficult to realistically assess what music they like to listen to and in what proportion. Flegal et al. (2019) show that some people struggle to estimate their own weight, and we can imagine that there might be a gap between declared preferences and the music that respondents actually listen to. For instance, it is possible that people tend to overstate listening to some more socially appealing music genres, and neglect to mention the less socially accepted music they like.

On the other hand, recommender systems mostly rely on observable data, often collected as traces of activity in online platforms. The huge amount of collected data should allow a good understanding of the users’ listening practices, and even though the context or sociodemographics are not explicitly collected, the data could be used to deduce some implicit information. For example, Way et al. (2019) estimate the relocation of certain users by analysing the changes in their IP address. However, the collected traces are often ambiguous and considered as implicit markers of preference (or negative markers, in the case of skipped songs for example) (Oard and Kim, 1998; Majumdar et al., 2009).

A way to have a complete understanding of people’s preferences would be to cross observable and declared data. This idea has been recently proposed (Cura et al., 2022) in the form of “augmented interviews” leveraging digital traces to inform and assist social science researchers conducting interviews.

2.2 Music representation

In order to quantify musical taste, one must first be able to segment the musical space itself. For this, music preferences can be assessed either directly using music items, like artists or songs (Bourdieu, 1984), or through the mediation of aggregated categories. In surveys, for the sake of brevity, preferences are often collected via set of music genres (Peterson, 1992; Bryson, 1996; Coulangeon, 2017). Even though representing music through genres may seem obvious, it is important to keep in mind that no universal genre taxonomy exists, thus using genres to depict people’s musical taste can create bias (Sordo et al., 2008). Music can also be classified by so called “mood”, that can be identified either through audio features (Soleymani et al., 2015; Delbouys et al., 2018) or through declared data (Rentfrow and Gosling, 2003). Bogdanov et al. (2013) used audio features in order to depict people’s musical taste.

2.3 Distinction and essence

In the literature, taste is often defined as a set of traits that distinguishes us from others and marks our individuality. In sociology, musical taste has long been studied in association with social class belonging. Bourdieu (1984), Peterson (1992), Bryson (1996) and Coulangeon (2017) show the connection between musical preferences and social class – people present their taste as a mark of belonging to their “in-group” while differentiating themselves from an “out-group”. Similar conclusions have been found in psychological studies, like Hargreaves et al. (2006), who studied adolescents and how they use music to build their social identity. Later, Lahire (2008) studies intra-individual behavioral variations and emphasizes that most people have preferences that are not typical for their social group, and thus taste is an individual characteristic. The need people have for distinctiveness or “uniqueness” in order to self-identify has been studied in psychology as well (Fromkin and Snyder, 1980).

The notion of distinctive identity is also reminiscent of that of digital identifiability, that is, to what extent people’s behavior (and digital traces of it) can be used to uniquely identify individuals. For example, De Montjoye et al. (2013) use mobility data and show that four spatio-temporal points are enough to uniquely identify 95% of individuals. Narayanan and Shmatikov (2008) use the Netflix Prize dataset to de-anonimize users through the movies they have watched on the platform. They show that 5–10 movie ratings are enough to identify most users. These studies present an extreme form of distinction, where each individual is literally identified in a unique way among all others. However, no such experiment has been run on music streaming data.

An alternate definition of musical taste would be a set of factors that characterize listening behavior of an individual. This is typically the definition implicitly adopted as the core principle for designing recommender systems, where the goal is to understand the essence of the user’s preferences in order to suggest them similar music. Two main approaches exist in recommender systems. In collaborative filtering, the idea is to assign users a descriptive vector, or embedding, based on similarities between other users. The same process is applied to determine the similarity between items. This can be done through matrix factorization (Koren et al., 2009) based on either implicit feedback, like streaming activity logs, or explicit feedback such as users’ collections of favorite items and playlisting of songs. Content-based recommendation, on the other hand, tries to define the items’ features that a user will respond positively to. These features can be represented by various tags (Pazzani and Billsus, 2007) that can be automatically computed based on audio features in the case of music (Cano et al., 2005; Van den Oord et al., 2013; Schedl et al., 2015) or social tags that can furnished by music providers or collected from the Web (Eck et al., 2007).

As concluding remarks, one may point out that the literature is rich with attempts to characterize musical taste, but they seem to be hard to reconcile, as they diverge on several key aspects. The first one is quantization of the musical space, the second being the data collected and the analysis methods. But most importantly there are conflicting hypothesis on the very nature of an individual’s musical taste. While social sciences emphasize the importance it bears in the construct of one’s self-identity, the emerging field of recommender systems assumes a form of homogeneity, even predictability of one’s taste.

This raises a series of open questions: to what extent is it possible to identify people based on their musical preferences? Assuming there are distinctive traits in one’s musical consumption, are these truly reflective of their global behavior?

3.1 Overview

For this study, we work with data obtained from the music streaming service Deezer,¹ that currently counts about 16M active subscribing users worldwide and has a catalog of 90M tracks. First, we collected explicit feedback data (i.e. “likes”) from 1M randomly selected users, who have been active during October 2022. Let us call this data sample D_L. Users can explicitly “like” songs, albums, and artists which then appear in their “favorites” collection. As of the date of the data collection, among these 1M users 87.1% of them had explicitly liked at least one artist, and 88.9% had liked at least one song. All together the users had liked 586 512 artists and 10 822 633 unique songs.

3.2 Distribution of musical items by received likes

The distribution of these items according to the number of unique users who like them follows a heavy-tailed distribution (Figure 1, top). For artists, the median value is equal to one — which means that at least half of them have been liked by only one user — while the average is around 38. The most popular artist has been liked by 86 877 users. We can thus see a huge disparity between the artists, with a few extremely popular artists that attract most of the users, and many artists that are almost unknown. The songs follow a similar popularity distribution, with a median of 1, an average around 18, and a maximum of 75 453 likes.

Figure 1

3.3 Distribution of users according to the size of their favorites’ collection

The distribution of users according to the number of artists they have liked similarly follows a heavy-tailed distribution (Figure 1, bottom). Half of the users have liked 10 artists or less, with an average of 26 liked artists per user. Some outliers exist, such as one user who has liked 7 497 artists. Users tend to like songs more than artists, with 215 favorite songs by user on average. The user experience on the platform contributes to this gap between explicitly liking artists and songs: indeed, the like button can be easily hit on a song while the user is listening to it, while liking an artist requires the user to specifically go to the artist’s page.

3.4 Item popularity metric

In our experiments, we will need to consider items’ popularity, and we found it would be easier to represent it with a discrete variable. We decided to split items into popularity bins, from the least to the most popular, in a way such that in each bin the sum of likes received by all items is the same. We arbitrarily fixed the number of bins to 6. Table 1 shows the distribution of the number of artists in each bin, as well as the maximum and minimum number of fans for artists in the bin.

3.5 Genre tags

Internally, Deezer uses a taxonomy of 33 main genres to classify music, and attributes one main genre tag to most musical items in the catalog. These tags are mostly provided by music labels and recording providers, but can also be manually annotated by human editors. According to main genre tags, rock, hip-hop, pop and electronic music are the most popular music genres among the users in our dataset (Figure 2).

Figure 2

3.6 Streaming data

In RQ3, we want to compare the users’ fingerprints calculated from their favorite items with those calculated on their streaming activity. To do so, we also use a separate data sample, D_S, containing 1 year of streaming logs from April 1st 2022 to March 31st 2023 (D_{S_year}) and favorite artists (D_{S_favart}) for 60K active users. We made sure that all users were active during the entire year, in order to make comparable sub-samples for a day (D_{S_day}), week (D_{S_week}), and month (D_{S_month}) with the same users in each subset.

3.7 Opening the dataset

Unless a user configured otherwise, the artists and songs that they have clicked as “liked” are publicly visible on the website using the user’s ID, and can be retrieved thanks to the streaming service API.² We personally did not use the Deezer API and got anonymized data directly from Deezer, containing both private and public users. In section 4, we show that some users can be identified in a unique way through their favorite items. Sharing the dataset could thus raise some serious privacy concerns and we have decided not to do it in this form. Further work on means to effectively anonymize this data is required. For example, Cormode et al. (2008) obtained promising results for anonymization of sparse bipartite graphs, which is exactly the structure of our data, and it would be interesting to consider how such anonymization methods would impact our experimental results.

4.1 Problem definition

Previous work in the cultural sociology literature (Lahire, 2008) has focused on musical taste uniqueness and individuality. Adopting this standpoint, we wonder if it is possible to find for each user a subset of their liked or streamed items, a fingerprint, that could be assigned to them only – that is, that would make them unique in the crowd. This raises several questions, that include: how many items need to be selected for each user to discriminate him/her from all the others? Are certain music genres more discriminative than others?

4.2 Methods

As already mentioned, the constraint of finding fingerprints of minimal size makes the problem hard to solve, and no method exists to do it in a reasonable time. Therefore, we first propose a baseline method that matches only the constraints of non-inclusiveness, and then present an approximate method to minimize the fingerprints’ sizes. We compute fingerprints based on favorite songs and artists on the 1M-user dataset, as well as favorite artists and streamed artists on a day, week, month and year time period on the 50K-user dataset. In the case of streams, we consider any user-artist interaction only once, no matter the number of times the user has streamed the artist.

4.2.2 Minimizing fingerprints’ size

The random sampling method is simple, but it likely creates fingerprints that are larger than necessary. In order to minimize the sizes of the fingerprints, we propose a greedy approach. Let G(U, I; L) be the user-item bipartite graph, where U is the set of vertices representing the users, I is the set of vertices representing the items, and L are the edges linking users and items: there is an edge (u, i) ∈ L if the user u has liked the item i. For a vertex u in U, V(u) are the vertices in I that are connected with u by an edge. For each item i, let W(i) be the set of users connected to i, and d(i) = |W(i)| its degree.

For a user u, we first compute the weights of each item in V(u), or, in other words the number of users that have liked each item in V(u). Then, the item i_min with the smallest weight is selected and appended to F(u). Then all the users that have not liked i_min are removed from the graph, as well as the item i_min, and the weights of the remaining items in V(u) are recalculated. The steps are repeated while there are other users than u remaining and |F(u)| < |V(u)|. The full algorithm, called F_{uniq_minsize}, is given in Algorithm 1.

Algorithm 1

We assume that, depending on the size of the dataset, the number of uniquely identifiable users will not be the same, and the same goes for the average fingerprint size. As the complexity of our algorithm is O(n*m), the computation time will be strongly impacted by the number of users in the dataset, as well as the number of musical items they have liked, which makes it complicated to run on huge datasets, like the whole population of a streaming platform for example. In order to estimate how the number of identifiable users and their fingerprint sizes evolve with the dataset size and the two algorithms, F_{uniq_rand} and F_{uniq_minsize}, we run both algorithms on subsets of 10ⁿ users of D_L, with n going from 3 to 6, and for each n we repeat the procedure on 10⁶/n different random subsets.

Also, we want to see the impact of the streaming period on those metrics, so we separately run F_{uniq_minsize} on D_{S_day}, D_{S_week}, D_{S_month} and D_{S_year}, and, additionally, D_{S_favart}.

4.3 Results

4.3.1 Users’ identifiability

To answer RQ1, we took interest in the number of users who are identifiable through their online activity. In the following sections, we will denote D_{L_uniq} the subset of D_L that contains uniquely identifiable users. As expected, songs seem to be more discriminative than artists: in D_L, 60% of users can be identified by their favorite artists, and 90% by their favorite songs.

However, users differ according to the number of items they have liked: the fewer favorite items users have, the harder they will be to identify (Figure 3). For instance, only 15% of the users with 5 favorite artists or less can be identified, while users who have liked more than 25 artists can be identified more than 95% of the time. The more items a user has liked, the more they become a so-called “power-user”, i.e. a user whose collection of items fully contains all the favorite items of other users who have smaller collections (Figure 4). In a dataset of 1M users, a user who has liked one thousand or more artists covers, on average, the favorite artists of more than 1% of all the users. Overall, users with at least one hundred favorite artists cover the likes of 41% of the users from the dataset, and users with more than one thousand favorite artists cover the likes of 32% of the users (Figure 5). However, “power-users” of different ranges mostly cover the same users. For instance, 93% of the users covered by users with 1000+ liked artists are also covered by users with 100–1000 liked artists, and 86% of the users covered by users with 1000+ liked artists are also covered by users with 100–250 liked artists. Therefore, the size of the dataset is a much more important factor for identifiability of the dataset than so-called “power-users”.

Figure 3

Figure 4

Figure 5

Additionally, we computed F_{uniq_minsize} on D_S. Expectedly, streams allow a much higher identifiability than likes, as users like much fewer artists and songs than they stream (Figure 6). Extending the time period for retaining stream logs strongly increases identifiability: one month of stream logs is enough to identify 95% of the users.

Figure 6

4.3.2 Fingerprint size

To answer RQ2, we first looked at the size of the assigned fingerprints. In D_L, we find unique fingerprints of an average size of 6.7 artists and 3.6 songs by drawing random items (Figure 7). For songs, the maximum size fingerprint is huge (176 songs to discriminate one user). Indeed, the dataset contains a few users with huge collections of liked items, up to almost 10⁵ favorite songs. The favorite items of such users are most likely to cover a lot of other users’ collections, which is why we would need this many items to discriminate them from others. However, considering the average and the median fingerprint size, which is 3 (for songs), we can assume that such a high fingerprint size is more of an exception than a rule.

Figure 7

With F_{uniq_minsize}, we find unique fingerprints of an average size of 2.3 artists and 1.4 songs. Among 1M users, 45% of them are identifiable with only one song. Table 2 shows that the average size of fingerprints based on songs increases only slightly with the size of the dataset. It can thus be assumed that even though the number of identifiable users will decrease in a larger dataset (Table 2), the average size of unique minimum size fingerprints based on songs will remain around 1.5.

Table 2

Distributions of fingerprint sizes, computed with F_{uniq_rand} and F_{uniq_minsize} based on favorite artists and songs, for different numbers of users in the dataset.

Artists
Sampling method	Number of users	Unique users (%)	Min F(u) size	Max F(u) size	Median F(u) size	Mean F(u) size	Standard deviation
Funiq_rand	1000	87.3	1	13	2	2.4	1.4
	10000	77.5	1	33	3	3.5	2.3
100000	67.7	1	58	4	4.9	3.6
871248	58.1	1	137	5	6.7	5.3
Funiq_minsize	1000	87.3	1	4	1	1.3	0.5
	10000	77.5	1	7	1	1.6	0.7
100000	67.7	1	10	2	1.9	1.0
871248	58.1	1	14	2	2.3	1.2
Songs
Funiq_rand	1000	96.8	1	8	1.9	1.7	0.8
	10000	94.4	1	33	2	2.2	1.2
100000	92	1	98	3	2.9	1.7
889017	89.9	1	176	3	3.6	2.4
Funiq_minsize	1000	96.8	1	2	1	1.0	0.1
	10000	94.4	1	5	1	1.1	0.3
100000	92	1	8	1	1.3	0.5
889017	89.9	1	194	1	1.4	1.1

The fingerprints’ size based on favorite artists D_{S_favart} (average 1.9, median 2) is comparable to one day of streams for the same users D_{S_day} (average 1.8, median 2), and slightly decreases with larger time periods (average 1.4, median 1 for a year of streams D_{S_year}).

4.3.3 Composition of the fingerprints

Another metric of interest to answer RQ2 is the fingerprints’ content. First, we compare the artists found in the fingerprints based on likes and streams, respectively D_{S_favart} and D_{S_year}. To this extent, we divide, for each user, the number of common artists by the total number of unique artists in both fingerprints. The found average ratio is around 1%, which means that there is no redundancy between the two kinds of fingerprints. Therefore, in a situation where anonymized streaming logs are shared, crossing this data with open access likes data should not lead to deanonymization, at least with the F_{uniq_minsize} method.

To have a deeper understanding of what kind of music is more discriminative, we compare the popularity and genre distributions of the fingerprint items with the users’ favorite items in general (on D_L). Unsurprisingly, the popularity of an artist or a song is an important indicator of whether or not it might be included in one’s fingerprint (Figure 8): the less popular the item, the more discriminative it is. As for the genres, the most popular ones, such as hip-hop, pop, rock and electronic music, seem to be underrepresented, while other, less popular genres, are overrepresented in the fingerprints (Figure 9).

Figure 8

Figure 9

If we consider musical taste through individuality and uniqueness, as Lahire (2008) did, we are then able to create fingerprints of musical taste. However, is individuality on its own a sufficient definition of taste, and do these unique fingerprints capture the essence of the users’ preferences?

5.1 Problem definition

We formulate the problem in a way similar to recommendation: a subset F(u) is considered as representative of V(u) if there exists a method F^⋆ such that ∀ u ∈ U, V(u) ≈ F^⋆(F(u)),. In other words, we consider that a fingerprint is representative if a method that can recover the initial set of items from it exists. Building an F^⋆ function is a ubiquitous task in recommender system research, where the problem is very similarly defined.

We chose to define F^⋆ as a simple prediction function based on the nearest neighbor algorithm, computing the proximity between the artists using matrix factorisation (Koren et al., 2009). For favorite items, we start by building a sparse artist-user matrix M, where M[u,i]=1 if the user u has liked the artist i. For streams, M[u,i]=1 if the user u has streamed the artist i at least once during the given time span. We then compute a singular value decomposition (SVD), and use the first 128 dimensions of the SVD as our artists’ embeddings. The artists’ nearest neighbours are then computed based on the Euclidean distances between their embeddings.

Let N(i) be a list of i’s nearest neighbors ordered from the closest to the furthest. Let w_i be a weight associated to each item i in a fingerprint F(u). This weight represents the number of items we need to recover from u. If all items in F(u) are equally important, then we want to recover the same number of items from each item in F(u): ∀ i ∈ F(U), w_i = (|V(u)| – |F(u)|)/|F(u)|. For a user u, F^⋆ returns a set of predicted items P(u) by simply taking, for each item i in F(u), the w_i closest neighbors of i from the list of i’s 150 most similar artists.

5.2 Evaluation proxy

The representativeness score of a fingerprint is calculated based on how close the predicted items are to the user’s favorite items. We propose two methods to compare P(u) and V(u):

• Item-wise. This evaluation is the most strict. The predicted items P(u) are compared exactly to the actual user’s favorite items (except the ones included in the fingerprint). The prediction accuracy for a user u is thus equal to |P(u) ∩ (V(u)|F(u))|/|P(u)|. This metric is widely used in recommender systems for offline evaluation tasks, where ground truth user-item interactions are available.

• Genre-wise. Here, we compare if the predicted items follow similar distributions in terms of genre as the items from the user’s actual favorite items. The prediction score is thus simply the L₁ distance between the distribution of genres in P(u) and the one in V(u).

Other metrics could also be used, based on the mainstreamness of the artists for example.

5.3 Experiments

5.3.2 Uniqueness vs essence

To answer RQ3, we now want to confront our two sampling methods, F_{uniq_minsize} and F_{rep_kmedoid}.

First, we run both sampling methods on uniquely identifiable users D_{L_uniq}, then run the prediction on both obtained fingerprints: the prediction accuracy from the unique fingerprints is lower (3% with item-to-item evaluation, 42% genre-wise) than the representative fingerprints (10% with item-to-item evaluation, 50% genre-wise) (Figure 10), meaning that the most discriminative items in the users’ libraries are not representative of their overall preferences.

Figure 10

Second, we found that only 279 435 users remain identifiable from D_L’s representative fingerprints, comparing to 507 037 in D_L overall. Thus, extracting the essence of one’s musical library most likely leads to a loss of the information that makes them unique.

Trying to quantify the essence and the uniqueness of one’s musical taste seems to represent two diverging goals, which require distinct computation methods.

6.1 Conclusion

Building on a large set of literature, we emphasize how preference elicitation encompasses several conflicting definitions. We propose to make two of them explicit, stressing constraints of uniqueness (respectively representativeness) as optimization goals adapted to distinction (respectively characterization) of an individual’s taste fingerprint. We show that these different constraints lead to diverging solutions which in turn suggests that scientific work addressing musical taste should probably reflect on their exact objectives and make their understanding of the term explicit.

We run our experiments using data from a major streaming platform, containing both explicitly liked content and streaming logs. In a first section of experiments, we show that in a sample of 1M active users, 90% can be identified by their favorite songs, and one or two songs is enough to identify 45% of the users. On another sample of 50K users, we also show that streaming logs are even more identifying, especially if collected for a long period of time – up to 97% of the users are identifiable via the artists they streamed for a year (RQ1).

However, the artists allowing to identify users are not the same when it comes to what they have liked or streamed. Also the more identifying items are expectedly the less popular ones, and by consequence, those from less popular genres (RQ2).

In a second section of experiments we propose a method to depict users’ preferences by creating representative subsets of users’ favorite items that we call fingerprints. This method can further be used in situations when concise information about the users’ preferences is needed: in recommendation systems, or scientific work that uses the concept of musical taste.

We show that the best items for identifying users are not the most representative of their preferences: using a prediction task, we can recover an average of 10% of the users’ favorite artists from the representative fingerprints against 3% from the unique fingerprints. Complementarily, only 279 435 users remain identifiable based on their representative fingerprints, against 507 037 in the initial set. It thus seems that the essence and uniqueness of musical taste are opposite concepts (RQ3).

6.2 Limitations

The experiments proposed in this work are nonetheless limited by the nature of the data used to conduct them. As we have emphasized, observable data are handy to collect at scale, but arguably they are non-perfect proxies of an individual’s true preferences. In particular, the information of explicit distaste is missing, though it appears to be a highly relevant indicator. An intuitive approach would be to leverage implicit feedback such as skips, but these are even noisier signals.

A more promising approach would be to build a richer, multi-modal dataset, containing both declared and observed data for a sufficient number of individuals. This will be the focus of our future work. Additionally, the evaluation of the fingerprinting methods could also be improved, in particular by means of an experiment involving the users themselves, for instance using an interface such as the one presented by Cura et al. (2022).

6.3 Addressing privacy issues

Unlike streaming logs, information about users’ likes is publicly accessible on the Deezer platform and most of their competitors, unless users specifically indicate their account as private. The fact that most users can be identified by their likes basically shows that a significant share of them are by default 1-anonymous Sweeney (2002), thus not anonymous. It reveals an important privacy issue – the usual practice of hashing the users’ IDs does not seem to be enough to anonymize a dataset. It can be especially compromising to share personal data, such as geolocation for example, combined with information about the users’ likes. Future work could be done to explore ways to aggregate or obfuscate such data in order to ensure k-anonymity, while keeping its expressiveness at the same time.

In the music information systems used on platforms, which must remain expressive for users, there is no category for describing music that would be both more precise than music genres (which are ill-defined categories) and more aggregated than the precise catalog items consumed by users: tracks, artists, and albums. Consequently, for lack of a better alternative, our results suggest that publicly available information about individuals’ music preferences should likely be aggregated at the level of music genres to strengthen anonymity (e.g. possibly defined as clusters of artists, whose size should be adjusted to ensure k-anonymity, with artist cluster sizes depending on k).