Taste or Addiction?: Using Play Logs to Infer Song Selection Motivation
1. Taste or Addiction?:
Using Play Logs to Infer Song Selection Motivation
Kosetsu Tsukuda and Masataka Goto
National Institute of Advanced Industrial Science and Technology (AIST), Japan
PAKDD 2017
2017/05/26
2. 2Play logs of online music services
Online music service
Play logs of Last.fm
User ID Artist Song title Timestamp
55 Madonna Crazy For You 1/5/2017 11:24:35
141 The Beatles Let It Be 1/5/2017 11:28:16
55 Lady Gaga Poker Face 1/5/2017 11:28:30
⋮
3. 3User’s song selection motivation
𝑡𝑡
By extracting a user’s logs, his play history can be generated:
it is observable when he listens to music and what he listens to
It is not observable why the user played each song
I played this song because ⋯
Research goal
Given a user’s play history,
infer the user’s song selection motivation of each played song
4. 4Factor 1: user’s usual taste in music
𝑡𝑡
User has his/her own taste in music
Hypothesize that such taste affects use’s song selection
I played these songs because I like rock music.
rockrockrock
5. 5Factor 2: user’s addiction to an artist
𝑡𝑡
User can be addicted to an artist regardless of his taste in music
(he watched the artist on TV or his friend recommended the artist)
Hypothesize that such addiction affects use’s song selection
My friend recommended an album of Lady Gaga.
I played these songs because I became addicted to her.
L.G. L.G.
6. 6What can our model do?
𝑡𝑡
Taste Addiction Taste
Taste: 73%
Addiction:
27%
Taste/Addiction ratio Taste
Rock: 45%
Hip-hop:
19%
Addiction
Lady Gaga:
45%Madonna:
32%
Our model can estimate user’s song selection motivation and
user’s statistics (e.g., taste in music and addicted artists)
7. 7What can our model do?
Our model can reveal users’ addiction-based music listening behavior
Users have high addiction degree in the morning
Users have low addiction degree on the weekend
𝑡𝑡
Taste Addiction Taste
𝑡𝑡
Addiction Taste
𝑡𝑡
Taste Addiction Addiction
⋮
9. 9Overview
Estimate user’s song selection motivation by using a topic model
Our model builds on the one proposed by Zheleve et al.
in WWW’10 that considers only user’s taste in music
Model by Zheleve et al.
Taste
Our model
Taste Addiction+
10. 10Session
𝑡𝑡
Divide user’s play logs into sessions
If the time gap between two songs is less than 30 min,
they belong to the same session
41 min 135 min
Session 1 Session 2 Session 3
𝑡𝑡
𝑡𝑡
11. 11Song generative process by Zheleve et al.
𝑡𝑡
User has a topic distribution that represents the user’s taste
One topic is assigned to each session based on the user’s taste
Artist is selected based on the artist distribution of the topic
Topic distribution = User’s taste
Prob
Topic 1 2 3 4 5
⋯
Topic 5
Artist distribution of topic 5
Prob
⋯
Eminem
JayZ
J.Cole
ICECUBE
2Pac
Eminem
J.Cole
Eminem
Topic 5 Topic 2
JayZ
Annekei
Annekei
12. 12Graphical model
Topic distribution = User’s taste
Prob
Topic 1 2 3 4 5
⋯
Artist distribution of topic 5
Prob
⋯
Eminem
JayZ
J.Cole
ICECUBE
2Pac
13. 13Our Model
User has a taste-addiction weight that affects his song selection
In an addiction mode, the user directly chooses a song
User has an artist distribution that represents the user’s addiction
Prob
Taste
Addiction
Taste-addiction weightAddicted artist distribution
Prob
⋯
LadyGaga
DianaRoss
BrunoMars
Adele
LMFAO
14. 14Song generative process in our model
For each session
Choose a topic
For each song in the session
Choose a mode (taste or addiction)
If taste mode, choose an artist according to the topic
Else if addiction mode, choose an artist
from addiction artist distribution
𝑡𝑡
Topic 5
Eminem
LadyGaga
LadyGaga
Taste
Addiction
Addiction
Addiction
Taste
Addicted artist distribution
Prob
⋯
LadyGaga
DianaRoss
BrunoMars
Adele
LMFAO
Adele
J.Cole
17. 17Research Question
Is adopting two factors, which are users’ taste in music and
addiction to artists, effective to model music listening behavior?
VS
Model by Zheleve et al.
Taste
Our model
Taste Addiction+
18. 18Dataset 1
Play logs of a music download service in Japan
Consists of 10 weeks logs from 1/1/2016 to 10/3/2016
8 weeks are used as training data and 2 weeks as test data
Users Artists Logs for train Sessions for train Logs for test Sessions for test
13,986 6,431 331,437 82,427 57,126 13,516
⋮
𝑡𝑡
𝑡𝑡
𝑡𝑡
8 weeks: training 2 weeks: test
1/1/2016 25/2/2016 10/3/2016
13,986 users
19. 19Dataset 2
Users Artists Logs for train Sessions for train Logs for test Sessions for test
2,850 12,360 872,614 106,840 201,966 24,958
Publicly available Play logs of Last.fm by Schedl in ICMR’16
Consists of 10 weeks logs from 1/1/2013 to 11/3/2013
8 weeks are used as training data and 2 weeks as test data
⋮
𝑡𝑡
𝑡𝑡
𝑡𝑡
8 weeks: training 2 weeks: test
1/1/2013 26/2/2013 11/3/2013
2,850 users
21. 21Results
Japanese service
Perplexity
Number of topics (K)
0
100
200
300
400
500
0 100 200 300
Zheleve
Ours
Number of topics (K)
500
800
1100
1400
1700
2000
2300
2600
0 100 200 300
Zheleve
Ours
Last.fm
Our model outperforms the model by Zheleve regardless of
the number of topics in both datasets
We did not evaluate for 𝐾𝐾 > 300 because if 𝑲𝑲 becomes too
large, many similar topics appear and it is difficult to carry
out the qualitative evaluation
23. 23Data
Use 8 weeks logs of a Japanese service with 𝐾𝐾 = 30
Compute parameters for the evaluation
8 weeks logs of
a Japanese service
Parameters
Users’ addiction ratio distribution
Per-hour addiction ratio
Per-topic addiction ratio
𝜃𝜃𝑢𝑢𝑢𝑢 =
𝑅𝑅𝑢𝑢𝑢𝑢 + 𝛼𝛼
𝑅𝑅𝑢𝑢 + 𝛼𝛼𝐾𝐾
𝜙𝜙𝑘𝑘𝑘𝑘 =
𝑁𝑁𝑘𝑘𝑘𝑘 + 𝛽𝛽
𝑁𝑁𝑘𝑘 + 𝛽𝛽 𝐴𝐴
𝜓𝜓𝑢𝑢𝑢𝑢 =
𝑁𝑁𝑢𝑢𝑢𝑢𝑢 + 𝛾𝛾
𝑁𝑁𝑢𝑢𝑢 + 𝛾𝛾 𝐴𝐴
𝜆𝜆𝑢𝑢𝑢 =
𝑁𝑁𝑢𝑢𝑢 + 𝜌𝜌
𝑁𝑁𝑢𝑢 + 2𝜌𝜌
𝜆𝜆𝑢𝑢𝑢 =
𝑁𝑁𝑢𝑢𝑢 + 𝜌𝜌
𝑁𝑁𝑢𝑢 + 2𝜌𝜌
24. 24Users’ addiction ratio distribution
Is the addiction ratio different from one user to another?
25. 25Users’ addiction ratio distribution
Users with > 0.9 addiction ratio achieve the second highest peak
Many users are addicted to artists to a certain extent: this result
further indicates the usefulness of considering the addiction mode
Prob
T A
0.72
Prob
T A
0.11
Prob
T A
0.53
⋮
13,986 users
0
1500
3000
4500
6000
7500
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frequencyofusers
Addiction ratio
T: taste
A: addiction
27. 0.55
0.6
0.65
0.7
0.75
0 2 4 6 8 10 12 14 16 18 20 22
27Per-hour addiction ratio
All users’ play logs between 6 AM and 7 AM
𝑃𝑃 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 0.22
𝑃𝑃 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 0.78
𝑃𝑃 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 0.59
𝑃𝑃 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 0.41
𝑃𝑃 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 0.75
𝑃𝑃 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 0.25
⋮
Sum 𝑝𝑝(𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡) and 𝑝𝑝(𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎) of all logs respectively and normalize their sum to 1
Hour
Addictionratio
Addiction ratio is high in the early morning and low at night
This knowledge can be applied to music recommendation
(e.g., recommend unknown artists to the user at night)
High addiction
28. 28Day of week based addiction ratio
0.6
0.61
0.62
0.63
0.64
0.65
Mon Tue Wed Thu Fri Sat Sun
Addictionratio
Addiction ratio is high on weekdays and low on weekends
These results enable us to analyze people’s music listening
behavior from a new viewpoint
Low addiction
30. 30Per-topic addiction ratio
All users’ play logs of top 20 artists in topic 19
Topic ID
0%
20%
40%
60%
80%
100%
10 8 19 20 1 4 26 12 21 30 2 9 13 18 11 6 14 25 5 24 15 3 22 27 23 16 29 7 28 17
addiction taste
𝑃𝑃 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 0.66
𝑃𝑃 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 0.34
𝑃𝑃 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 0.52
𝑃𝑃 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 0.48
𝑃𝑃 𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 = 0.74
𝑃𝑃 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎 = 0.26
⋮
Topic 8 (low addiction) is related to anime songs and
topic 28 (high addiction) is related to old Japanese artists
These results would be useful to generate a playlist
31. 31Summary
2Carried our a quantitative evaluation
Our model outperformed the one considering only user’s taste in music
3Carried out qualitative evaluations
Users’ addiction ratio distribution, per-hour addiction ratio etc
1Proposed a new model for music listening behavior
Our model considers user’s usual taste in music and the addiction to artists
Contributions
Future work
Use the knowledge obtained from log analysis for
artist recommendation and play list generation
Extend our model by considering the time transition of addiction