Experiment Design
As we would like our experiment not to be biased we try to design a rather neutral way to use a Spotify account on a daily basis. It should enable us to see:
-
Whether our choices influence our recommendations (and if yes, how so)
-
Whether our recommendations are influenced by factors that are not related to our personal tastes (if yes, what the recommendations are)
-
Track as accurately as possible what happens on the platform as we listen to music
We should pick two songs that are as different as possible and see whether there is a convergence or divergence as far as music recommendations over time.
Accounts
We create 4 accounts:
-
Two “polar opposites” - account A & B (specific opposite music styles)
-
1 diverse account - account C ( diverse music styles without the influence of Spotify at a first point)
-
1 reference point - account O (account on which we only listen to mainstream music and then only rely on what Spotify suggests to us from the beginning to the end)
Objective
→ The objective is to evaluate whether:
-
there is a diversification of the music output for each account.
-
there is a convergence of these 3 different profiles towards the reference point.
Input Selection
For the 2 “polar opposite” accounts (A and B): We pick two songs with extremely different characteristics
-
Overall sound (raw audio)
-
Geographical origin (use world radio map)
-
Music genre
-
Time period
The website Every Noise at Once gives us access to a library of songs that are mapped out along two axes:
-
mechanical-organic axis (vertical)
-
atmospheric-bouncy axis (horizontal)
⇒ We can choose two songs that are “polar opposites” along those to axes.
According to the map, we choose the following songs to start streaming on account A and B:
-
Agneton - “Back to Hyperion - Original mix” (Goa Trance)
-
Manuel Cardoso - “Introitus” (Portuguese early Renaissance)
N.B: These songs are also interesting because they are relatively far away from mainstream music ( which is taken as a the reference point)
Protocols
Protocol for account A & B
Week 1 (Input only)
Day 1
-
Listen to song a*/b*
-
Save song a*/b*
-
Check whether they are included in an album
-
If yes, listen to 4 other songs in the album
-
Save each one of those songs
-
Day 2-7:
-
Save the song
-
Check whether it is included in an album
-
If yes, listen to 4 other songs in the album
-
Save each of those songs
-
We notice that the first music content that is offered to us when we open Spotify’s homepage is not related to our previous consumption (Sorties - with 10 preset playlists)
However, there is another category below that is related to our previous consumption (Plus du genre de and weekly recommendations…)
Week 3-4:
Each Day
-
Listen to 5 songs from the weekly recommendations
Protocol for account C
Week 1 (Input only)
Day 1
-
Open private mode on your browser
-
Listen to song a*b*c*d*(starting points)
-
Check whether they are included in an album
-
If yes, listen to 4 other songs in the album around each starting point
-
Save each of those songs
-
Day 2-7:
-
Check whether it is included in an album
-
If yes, listen to 4 other songs in the album
-
Protocol for account O
Week 1-4
Day 1-7
-
Check Spotify’s Global Top 200
-
Listen to the 5 first songs on the list
Map
To present the results, we decided to reproduce the four axes of the map of Every Noise at Once. That to maintain the analysis within the genre-space in which we started.
To reproduce the variables
We found that we another way to represent and measure the Organic-Mechanical axe would be the accousticness and to measure the Atmospheric-Bouncy the danceability of a song. To look at this variables we used an audio tracking page from Spotify that reveals the audio features of any song in Spotify.
​
We finally computed the scatter-plot with desmos.com , an online graph calculator.
We decided to use random samples and do a scatter plot, where the starting points are represented by an x , the imputs by full points and the outputs by circles.
