The Ethics of Spotify’s Machine Learning Algorithm for the Personalization of Song Recommendations

Shray Dewan

1

Data Ethics

01

02

03

04

Table of Contents

Executive Summary

Context & Problem

Ethical Analysis

Guidelines

2

2:00 min

6:00 min

6:00 min

6:00 min

Executive Summary

01

3

Data Ethics

1 Executive Summary

• Spotify has emerged as the music streaming giant due to its accurate personalized recommendation algorithm
• Two main actors: Spotify users and artists
• Amount of data collected and the way algorithm works → ethical problems
• Data privacy
• Altered interactions between users and artists

4

1 Executive Summary (cont’d)

• Recommendations
• 1. Provide an opt-out or opt-in option from data tracking
• 2. Offer more diverse song recommendations to users
• 3. Remove the monthly listeners feature from artists’ pages

5

Context & Problem

02

6

Data Ethics

2.1 Music Streaming Industry

• Transition to streaming services
• Music streaming market share breakdown
• Spotify: 500M monthly users - biggest platform

7

2.2 Spotify

• Known for its personalized playlists and accurate song recommendations

8

2.2 Spotify

• Employs artificial intelligence (machine learning algorithm) for its song recommendation algorithm
• Collects 1 TB of user data per day

9

2.3 Spotify’s Algorithm: Breakdown

• Called Bandits for Recommendations as Treatment (BaRT)
• Multi-armed bandit algorithm
• Objective: Recommend songs as similar to the songs a user likes
• 3 types of predictor variables:
• Collaborative Filtering (Exploitation)
• Collaborative Filtering (Exploration)
• Content-Based Filtering

10

2.3 Spotify’s Algorithm: Variables

11

2.3 Spotify’s Algorithm: Example

Original Playlist

Order by

12

2.3 Spotify’s Algorithm: Example

Original Playlist

Order by

13

(-60, 0)

(0,1)

(0,1)

(0,1)

(0,1)

(0,100)

(47, 53)

MED

(0.29, 0.75)

LOW-MED

(0.31, 0.71)

MED

(-14.30, -5.14)

LOW-MED

(0.11, 0.40)

LOW

(0.20, 0.43)

LOW

(72, 170)

MED

2.3 Spotify’s Algorithm: Example

Enhanced Songs

Order by

14

2.3 Spotify’s Algorithm: Example

Enhanced Songs: Why

Order by

15

2.3 Spotify’s Algorithm: Example

Enhanced Songs: Why

Order by

16

2.3 Spotify’s Algorithm: Example

Enhanced Songs: Why

Order by

17

2.4 Problems with Algorithm

1. Privacy of user data
1. Data security, third-party usage
2. Cold-start problem - new artists do not have user data
3. Amplification of biases in listening history

18

Ethical Analysis

03

19

Data Ethics

3.1 Privacy

• Tension: providing a personalized listening experience and individual privacy
• Q: How does Spotify handle user listening data, including storage, analysis, and selling?
• Storage of 1 TB/day/user
• Analysis of what and when user listens to
• Selling to third parties of user listening data

20

3.1 Privacy (cont’d)

• Extension: Target’s analysis of sales data to pinpoint pregnant customers

​

“Pole applied his program to every regular female shopper in Target’s national database and soon had a list of tens of thousands of women who were most likely pregnant. If they could entice those women or their husbands to visit Target and buy baby-related products, the company’s cue-routine-reward calculators could kick in and start pushing them to buy groceries, bathing suits, toys and clothing, as well.”¹

​

​

​

¹ Charles Duhigg, “How Companies Learn Your Secrets,” New York Times, February 16, 2012, https://www.nytimes.com/2012/02/19/magazine/shopping-habits.html.

21

3.1 Privacy (cont’d)

• Example: If you listen to the Calm Labor & Delivery playlist, Spotify may analyze your listening data and pinpoint you as someone who could be pregnant
• Spotify could then sell this information

22

3.2 Classification

• Q: How does Spotify’s algorithm resemble and fail to resemble a classification system?
• Classification system - “set of boxes (metaphorical or literal) into which things can be put to then do some kind of work;” three properties:
• 1. “There are consistent, unique classificatory principles in operation.”
• 2. “The categories are mutually exclusive.”
• 3. “The system is complete.”²

² Geoffrey C. Bowker, and Susan Leigh Star, “Introduction: To Classify is Human,” in Sorting Things Out: Classification and Its Consequences (The MIT Press, 2000), 10-1.

23

3.2 Classification (cont’d)

• No fixed categories but can think of an ordinal scale of categories based on similarity to user song preferences
• Example for classification of song:
• 7 - very similar
• 6 - similar
• 5 - somewhat similar
• 4 - neither similar nor different
• 3 - somewhat different
• 2 - different
• 1 - very different

24

3.2 Classification (cont’d)

• Does Spotify’s algorithm meet the principles of a classification system?
• 3. Complete system → Yes
• Ability to analyze every song on Spotify’s database
• 1. Consistent principles → No
• Cold-start problem: new artists’ songs are only based on content-based filtering, not collaborative filtering
• 2. Mutually exclusive → No
• Some songs can have elements of similarity and dissimilarity to user’s music taste

25

3.3 Music “Echo Chambers”

• Q: How does Spotify’s algorithm create music “echo chambers”?
• Echo chambers - “a social structure in which:
• 1. One must subscribe to a certain belief system to be a member.
• 2. That belief system includes the belief that all non-members are untrustworthy, and all members trustworthy.”³

​

³ C. Thi Nguyen, “How Twitter Gamifies Communication,” Forthcoming in Applied Epistemology, ed. Jennifer Lackey (OUP): 31.

26

3.3 Music “Echo Chambers” (cont’d)

• Algorithm only feeds users “very similar” songs (positive feedback loop)
• Easy for users to get stuck listening to the same music and thinking only their music is ‘good’ (Condition #1)
• Fosters belief that others without same music interest have ‘bad’ taste (Condition #2)

27

3.3 Music “Echo Chambers” (cont’d)

• Amplification of biases in “echo chambers”
• Listening to and having recommendations for similar music → increase biases in music niche

28

3.4 Gamification for Artists

• Q: How does Spotify’s algorithm gamify popularity for artists?
• Gamification - “increases our motivation by changing the nature of the activity”
• “Often, the goals of ordinary activity are rich and subtle.”
• “When we gamify these activities, we change those goals to make them artificially clear.”⁴

​

⁴ Nguyen, “How Twitter Gamifies Communication,” 3.

29

3.4 Gamification for Artists

• Artist popularity: traditionally difficult-to-measure value
• Spotify gamifies artist popularity with monthly listeners and artist rank

30

3.4 Gamification for Artists

• Monthly listeners: not the most accurate measure of artist popularity
• Based only on song listens this month
• Popularity: Ice Spice > J. Cole? Frank Ocean? Lorde?

31

3.4 Gamification for Artists

• Has led to value capture for artists
• Value capture - “occurs when:
• 1. Our natural values are rich, subtle, and hard-to-express.
• 2. We are placed in a social or institutional setting which presents simplified, typically quantified, versions of our values back to ourselves.
• 3. The simplified versions take over in our motivation and deliberation.”⁵

⁵ Nguyen, “How Twitter Gamifies Communication,” 21.

32

3.4 Gamification for Artists

• Artists have adopted new strategies to gain “popularity” (value capture)
• 1. Change in focus: albums → singles
• 2. Release singles more frequently
• 3. Release music that falls into pre-established music niches (music “echo chambers”)

33

3.5 Transparency over Accountability

• Q: Does Spotify prioritize transparency over accountability?
• Transparency of algorithm through releasing and discussing metrics
• Limit of transparency: “disconnected from power”
• “If transparency has no meaningful effects, then the idea of transparency can lose its purpose.”⁵
• Unethical practices continue despite ethical concerns
• Accountability necessary

​

⁵ Mike Ananny and Crawford, Kate, “Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability,” new media & society 20, no. 3 (2018): 978.

​

34

3.6 Ethical Questions Summary

1. How does Spotify handle user listening data, including storage, analysis, and selling?
2. How does Spotify’s algorithm resemble and fail to resemble a classification system?
3. How does Spotify’s algorithm create music “echo chambers”?
4. How does Spotify’s algorithm gamify popularity for artists?
5. Does Spotify prioritize transparency over accountability?

35

Guidelines

04

36

Data Ethics

4.1 Privacy

• Objective: improve data privacy for Spotify users
• Spotify should provide an opt-out, or better yet, an opt-in from data tracking
• Tradeoff: reduced personalization
• Would benefit users who do not want their data potentially stored, analyzed, and sold

37

4.2 Music “Echo Chambers”

• Objective: reduce frequency of music “echo chambers”
• Spotify should provide more unexpected, diverse song suggestions that are only somewhat related to user’s music preferences
• Consider employing if user reaches a music “echo chamber” threshold

38

4.3 Gamification for Artists

• Objective: reduce level of gamification for artists
• Spotify should hide the monthly listeners features
• Spotify should only have its artist rank feature, if anything
• Incorporate album success in rank

39

References

05

40

Data Ethics

5.1 References

Ananny, Mike, and Crawford, Kate. “Seeing without knowing: Limitations of the transparency ideal and its application to algorithmic accountability.” new media & society 20, no. 3 (2018): 973-89.

​

Bowker, Geoffrey C., and Susan Leigh Star. “Introduction: To Classify is Human.” In Sorting Things Out: Classification and Its Consequences, 1-28. The MIT Press, 2000.

​

Duhigg, Charles. “How Companies Learn Your Secrets.” New York Times, February 16, 2012. https://www.nytimes.com/2012/02/19/magazine/shopping-habits.html.

​

krn6. “The End of an Era: The Death of the Album and Its Unintended Effects.” Gnovis Journal, February 28, 2014.

​

Marius, Hucker. “Uncovering How the Spotify Algorithm Works.” Towards Data Science, November 23, 2021.

​

Michell, Tay. “Spotify's Algorithm: Helping Or Hurting Musicians?” Arts Management & Technology Laboratory, April 28, 2022.

​

Nguyen, C. Thi. “How Twitter Gamifies Communication.” Forthcoming in Applied Epistemology, ed. Jennifer Lackey (OUP): 1-40.

​

Platinga, Bo. “What do Spotify’s audio features tell us about this year’s Eurovision Song Contest?” Medium, April 28, 2018.

​

Spotify for Developers. “Get Track's Audio Analysis.” Accessed May 11, 2023. https://developer.spotify.com/documentation/web-api/reference/get-audio-analysis.

​

Tirumala, Sameer. “APIs Zero to Hero: ft. Spotify.” Medium, September 27, 2022.

​

Wall Street Journal, “Spotify Breaks Down the Mapping Tech Behind Its Algorithm.” April 17, 2023.

​

41