Every participant is asked to listen to a set of songs as the experiment begins. This first playlist consists of a very specific assortment of songs which are arranged in a specific order so as to elicit the participant's response. Note that for the purposes of the experiment, we limit the recommendations to the works of the top 10 artists in the world based on the listening activity observed on Deezer (i.e., the ten most popular artists on Deezer) during the years 2018-2019. Hence, for building this initial playlist, and all the remaining playlists in the experiment, we utilize the music profile of the top 10 artists such as Drake, Ed Sheeran, Coldplay, Imagine Dragons, XXXTENTACION, etc. We make this conscious choice of restricting our study to this set of 10 artists because of their popularity and the higher likelihood of participants to be already aware of their past works. Involving relatively unknown artists in our analysis may confound the eventual results due to some participants already being aware of their music and some not.
Through this first listening session, we simultaneously measure two kinds of behavioral tendencies of each participant: (1) the amount of variability she prefers in successive songs within a listening session and (2) the kind of music she prefers. To do so, we track how long a participant listens to a given song and whether and when she skips the song to move to the next song. Based on such information, we immediately determine (1) whether a participant prefers to seek various tastes or form habitual tastes ("variety-seeking" vs. "habit-forming") as well as (2) whether a participant prefers a certain genre over the other genre in a given playlist ("pop-rock" vs. "rap"). As such, each participant, as soon as the first listening session ends, gets to have two behavioral labels regarding her variability preference and genre preference, respectively. Depending on which treatment a participant is assigned to, only one of the two behavioral labels of that participant will be used. For those who are assigned to the control group, neither information will be used.
Before beginning the second session, the participants are randomly assigned to one of the three groups which decides the kind of recommendations that will be given to them. Accordingly, we have two treatment groups and one control group:
1. Treatment Group 1 is given a new set of songs based on an algorithm that follows the proposed recommendation methodology in our paper (of the same title). This recommendation is based on the amount of variability that a participant prefers during the listening session (variability preference). Hence, for a participant who is classified into this group, we already know whether she is more inclined to be 'habituated' or 'variety seeking.' We have two separate playlists that cater to each of these distinctive tastes which is then fed to the participant in the second session based on which suits her the best. These playlists are generated from the works of the top 10 artists by the algorithm that can estimate the artistic difference between two songs using the cosine dissimilarity index as shown in the paper.
2. Treatment Group 2 gets recommended songs using content-based collaborative filtering technique that is the status quo in the music streaming industry. Here, the recommendation is more concerned with tracking the kind of music, for example, genre, that the participant prefers (genre preference). The first session already allowed us to learn whether the participant likes a certain kind of music, here: rap (for example, music from Drake, XXXTENTACION, Eminem, etc.) and pop-rock (for example, music from Coldplay, Ed Sheeran, etc.). With this understanding, we then feed the participants in treatment group 2 a playlist that best reflects their musical taste between rap music and pop-rock music. These playlists are also randomly generated from the songs released by the top 10 artists by an algorithm that seeks to classify songs based on their genres (acoustic attributes).
3. The control group is given a new set of randomly selected songs. These songs are randomly selected from the options available out of the total songs released by the top10 artists considered in our study.
The group assignment between the first and the second listening sessions is the only intervention designed for the present study.
Note that the participants in this study are all equally compensated. Hence, if there are any participants who complete the experiment by clicking on the skip button too often (without really giving any song in the playlist a reasonable listening time), we plan to exclude them from the analyses.