Listening to music radios is an activity that since the 20th century is part of the cultural habits for people all over the world. While in the case of analog radios DJs are in charge of selecting the music to be broadcasted, nowadays recommender systems analyzing users' behaviours can automatically generate radios tailored to users' musical taste. Nonetheless, in both cases listening sessions do not depend on the listener choices, but on a set of external recommendations received. In this preliminary study, we propose a model for estimating features' variation during listening sessions, comparing diferent scenarios, namely analog radios, personalized and not-personalized streaming radios. In particular, we focus on the analysis of track popularity and semantic information, features well-established in the Music Information Retrieval literature. The presented model aims to quantify the possible impact of the sessions' variation on the user listening experience.
• General and reference → Evaluation; • Information systems → Recommender systems.
Since the first part of the 20th century, the radio has been one of the
main media thanks to which listeners can enjoy music. Its influence
on music consumption has been enormous, due to its capacity to
determine content difusion and diversity [
However, the use of personalized services in online spaces is
at the center of the scientific and public debate, due to its proven
tendency in fostering polarization in society [
In this work, we aim at characterizing the listening session
variations. We focus on two diferent aspects: popularity, feature
historically related to the "long tail" problem [
The paper is structured as follows. Section 2 provides an overview of previous work related to the research lines of this study. We then propose a methodology in Section 3, which includes a description of the model. Section 4 describes the case study examined, together with the results obtained. Finally, conclusions and future work are discussed in Section 5. 2
The limits of considering accuracy metrics as the yardstick for
evaluating recommender systems have been discussed in the last two
decades, since the problem of formulating user-centric measures
has been approached and new metrics beyond-accuracy have started
to be proposed [
In the field of MIR, the concept of diversity has been often
associated with aspects of musical tastes and listening habits [
Alongside, from diferent points of view, temporal dimensions of
the listening experience have been the object of study. For instance,
a dynamic model of the listening experience has been proven to
be efective for measuring the evolution of musical taste over time
[
In conclusion, the study of the relationships between
sequentially ordered objects is part of the sequence-aware recommender
system framework [
The main goal of our model is to characterize indicators of variations for a recommendation list. Specifically, we focus on music recommendations, creating two metrics, one for the track popularity, and another one for the semantic content represented by a set of tags describing the track music genre. 3.1
We define a recommendation list R as an ordered set of n tracks R = {r1, r2, ..., rn }, where each track ri has two diferent attributes: 1) a set of semantic tags si , where si = {taд1, taд2, ..., taдn }; 2) a popularity index pi , computed as the sum of the track and artist popularity values pi = popri + popar t ist (ri ). To compare lists generated in diferent contexts, we study how track attributes vary. In detail, we adopt two temporal strategies for a sequential analysis of the tracks: we first consider single tracks and we then form groups according to their temporal location. 3.1.1
Single Track Level. A music recommendation list can be considered as a set of songs potentially listened in sequence. Our first goal is to model the deviation of each track with respect to previous ones in terms of semantic information and popularity. Therefore, we define two recursive functions, which at each step return an indicator i containing information about these variations.
Let’s consider Sn as the union of all the tags observed until the track rn and in as the number of tags of the n-th track which have can be described with the following formula: S0 = s0 Sn = Ð s i=1
i i0(s) = 0 in(s) = |sn \Sn−1 | |sn | n where |S | represents the cardinality of a set S, and the \ operation between sets represents the relative complement i.e. objects that belong to the first set and not to the second. Considering that multiple tags can be associated with the same track, the relative complement is divided by the cardinality of sn , giving the same weight to every tag. As the space of tags is finite, when n → ∞ relationship can be imagined as a hypothetical session, where a user listens to a radio without ever stopping, and where eventually the variations in terms of semantic tags will become null at some s ⇒ in( ) → 0. This point.
In the case of popularity, at each track is associated a value p such
sum of powers of two with the previous track popularities:
that p ∈ [
⇒ in(p) → 0.
We observe that by the nature of this iteration, the n-th variation will be influenced by all the previous tracks popularity in the list, but giving more relevance to the closest tracks. In this case, the convergence to zero is not guaranteed when n tends to infinite. The only case when variations decrease until reaching zero is when the popularity attribute tends to a constant value, hence if pn → c for
Finally, for deriving a central tendency of the variations for each list R we compute the median of the obtained values:
I (s) = median(i(js=)1:n ), I (p) = median(i(jp=)1:n ),
R R
However, the proposed modelization represents a situation far from being real. Indeed, it is hard to imagine a person who turns on the radio and listens to the same station without stopping or changing a song. If we consider temporal relationships between tracks played in very distant time spans of a listening session, we may include information which is not relevant. In order to avoid that, we present in the next section an approach to estimate how a recommendation list can embed diferent degrees of variations, considering group of tracks instead of single ones. define sдi for the group as: 3.1.2
Group Track Level. Starting from a recommendation list R, we divide it into uniform groups of M tracks. As a first step, we have to define how the track attributes, semantic tags and popularity, are aggregated to create a group-level representation.
With respect to semantic information, we combine all the tags related to the tracks in the group. If дi = {ri∗M , ..., r(i+1)∗M −1}, we sдi = (i+1)∗M −1 Ø j=i∗M s j , ∀i = 0, 1, ...
(4) (1) (2) (3) track ID sions with diferent variation characteristics. (5) (6) (7) (8) According to the model described in Section 3.1, we compute indicators of variations using as input the recommendation lists presented in Section 4.1, and the results are presented in Figure 1 and 2. For both features, values are normalized between 0 and 1. Given the small sample size and the nature of the experiment, these results can only be considered as descriptive.
Figure 1 shows the variations i computed using the formulas (1), (2), (6), and (7). We only display the variations of personalized recommended lists generated starting from five seed tracks (B, D, F, H, J). In the case of semantic information, we notice how single level variations describe a decreasing trajectory, converging to zero, in line with the hypothesis discussed in Section 3.1.1. For the popularity feature, the variations tend to remain constant. In addition, we do not observe any diference between lists created from diferent seed tracks. Analyzing the variations at a group level, we see how removing the long-term dependencies the variations behave independently not following a specific tendency, but depending more consistently from the conformation of the groups. In this case, we observe similar behaviours for both features.
Figure 2 shows the obtained indicators described in formulas (3) and (8). In terms of semantic information, we observe that the indicators produced by the single level model describe sessions more homogeneous than in the case of the group level. Indeed, the values of the indicators are lower for individual tracks than the correspondent values obtained at a group level, showing how the long-term dependencies impact the variations, leading to a median value equal to zero in several cases. In particular, for the non-personalized lists, the median of the variations is zero in half of the cases. It can be seen as a consequence of extensive use of semantic relationships while building the recommendations, being not tailored to specific users. In terms of popularity, we observe that, when comparing tracks from the same decade (A-B, C-D, E-F, etc.), the lists created starting from the tracks with lower popularity (A, C, E, etc.) present more variations than the ones with greater popularity, both in personalized and non-personalized settings. These results partially reflect the impact of the well-known popularity bias in recommender systems, describing a situation where starting from a popular track, other popular tracks tend to be recommended, and variations are limited.
From the group level perspective, computing the Pearson correlation coeficient r we notice that there is a higher correlation between non-personalized and personalized lists than in the case of the single level model. Indeed, for single level indicators we have r=0.15 for the semantic information feature and r=0.44 for the popularity, while for the group level we have r=0.55 for semantic information while r=0.79 for popularity. Even if these results are sensitive to the group size, we imagine that joining diferent tracks’ features in a single group attribute, as defined in Section 3.1.2, can lead to a mitigation of variations, hence to obtain results more correlated between personalized and non-personalized systems. Furthermore, the removal of long-term dependencies in the group level model impacts the degree of variations observable, leading to higher median values for both features. In this case, no particular efects have been noticed according to the seed tracks considered.
In conclusion, apart from the case of group level semantic information, we observe that the overall indicators of analog radios result to have higher values on average than the personalized and nonpersonalized recommendation lists. However, the baseline build is weak, given the few numbers of analog radio considered, hence to make a proper comparison it is needed to increase the amount of data analyzed, as discussed in the following section. 5
We have presented a model for analysing the variations of two items’ features in a recommendation list: popularity and semantic information. In our experiment, we have compared several music recommendation lists generated in diferent contexts: personalized and non-personalized streaming radios, and analog radios.
Several limitations of the obtained results are related to the scarcity of data. Indeed, the lists taken into account cannot be considered as representative, and a larger amount of data is needed for validating empirically our model. In this preliminary stage, we do not have evidence about the advantages or disadvantages of using personalized or non-personalized systems for generating diverse listening sessions, but comparing against radio playlists generated by humans, we believe that diferences with algorithmicdriven listening experiences can emerge.
Moreover, being the details of the recommender systems’ architecture used in this study unknown, it is dificult to make comparisons between personalized and non-personalized lists, and also to link observed efects to a specific cause. For instance, the popularity bias may potentially be reinforced more in non-personalized popularity-based recommender systems than in personalized settings.
Finally, for properly evaluating the impact of personalization in music recommender systems, we need to include in the model additional information about the users’ tastes and listening behaviours. In the presented work, personalization has been analyzed as a de facto phenomena, but further work needs to be done for evaluating its relationship to the recommendation provided. As an alternative line of research, users’ perceptions of listening sessions’ variation can be considered, comparing them with the indicators designed in this work. 6
This work is partially supported by the European Commission under the TROMPA project (H2020 770376). The authors would like to thank Jacopo Temperini, Elena Congeduti, and the anonymous reviewers for their comments and feedback.