Abstract

 

The Music Consumption on Spotify project is an empirical study of Spotify using quantitative data taken from experiments as Spotify users which were developed into and analyzed in a website. The question treated in the project concerns how Spotify’s suggestion system affects music consumption. The hypothesis we sought to test were whether Spotify increased the number of songs streamed by users without expanding the variety of songs, whereas on a collective level whether the supply of music in the app was homogenized. The graphic analysis of the results demonstrated a trend that was not in favour of homogenization collectively, and pinpointed the effect of filter bubbles at the individual level as shown by a relative lack of variety.

 

Reflexive Analysis

 

The results of the experiment demonstrated that for accounts A, B, C and O, the output corresponded in terms of genge as well as the accoustiness and danceability value to the input regardless of the difference of input values across the four accounts. When plotted on the graph, this pattern could be seen as points of input and output stayed within the same quadrants of the graph. Likewise, the categorisation of input and output values for each song in terms of country of origin showed different degrees of variety depending on the account but did not greatly diverge from the main countries of origin for that account; there was always a large overlap between the country of origin of the input and output variables. One must also keep in mind that through streaming, one has access to extensive lists of music on demand, which helps to reinforce one of the original hypotheses of the experiment. Carried out over a period of four weeks, the experiment indicates that Spotify does not significantly increase the variety of the music choices recommended to listeners.

When looking solely at the output data of the experiment as it was plotted in the graphs along the axes of accustincness and danceability, the points did not show any signs of convergence with respect to all four accounts, as the output occupied a limited space focused in or around a single quadrant. Account O, the mainstream music oriented account, presents an exception as it had the greatest variety of both input and output variables of all four accounts, yet its output data much like that of account A diverges along the horizontal axis and does not begin to resemble any one of the other four accounts. Rather, the wide gamut of account O encompasses much of the areas of the other three accounts without them growing to resembling it from input to output data. Thus, collectively across the four accounts the supply of music in the  does not tend toward homogenisation, which contradicts the original expectations and shows that our experiment was successfully able to disprove second claim of the hypothesis.

The theory which one can develop from these results would thus be that Spotify affects the music consumption by targeting individual consumers as opposed to just large audiences, which was demonstrated by the data that tends to reinforce the first hypothesis, whereas the refutation of the claim that homogenisation of music supply occured collectively demonstrates the application’s ability to target niche audiences.

 

Limitations

Although the application of our methodology gave conclusive results to the question that we sought to answer and disproved on of our hypotheses, allowing us to establish a theory, it still faced limitations and potentially unaccounted variables. Our data may have been affected by the fact that accounts A and C were run on an iPad and computer respectively and the Spotify app in those devices has features similar to that of the premium version of the cellphone and may have had an effect on the variety of output. Likewise, the experiment was only run for five weeks in total including the trial period and thus had a limited temporal scope. If the time had been longer there may have been a more significant divergence in the data, hence out thesis is only applicable to a constrained time period. Furthermore, our results would have been more effective had an entire class run the experiment so that the data of 20 accounts instead of 4 could be cross referenced.