ASONAM2023_presection_slide_track-recommendation.pdf

Presented at the 2023 IEEE/ACM International Conference on
Advances in Social Networks Analysis and Mining.

1. An Effective Graph-based Music Recommendation
Algorithm for Automatic Playlist Continuation
Toshihiro Ito
College of Information Science
University Of Tsukuba
Ibaraki, Japan
[email protected]
1
Hiroaki Shiokawa
Center for Computational Science
University Of Tsukuba
Ibaraki, Japan
[email protected]

2. 2
Outline
1.What is APC?
2.Related Work
pContent-based
pCollaborative Filtering
3.Propose Method
4.Experiments

3. 1. What is APC?
3

4. 4
Automatic Playlist Continuation (APC)
APC is a major feature in Spotify, Apple Music,…etc.
!Sugar / Maroon 5
!See you again / Charlie Puth
!Attention / Charlie Puth
…
Playlist
Recommendation System
!Payphone / Maroon 5
!We don’t talk anymore
/ Charlie Puth
…
Results
Male Artists2010’s
Pop music

5. 5
What is the goal of our research?
Goal: Building a high-quality APC system
Challenges:
nComplexity of playlist features and user preferences
pBased on the user’s background (Age, Country, Relationships,…etc.)
nA vast number of tracks on a music streaming platform
pSpotify: over 100 million tracks

6. 6
Problem Definition
APC Problem
ℳ: set of music tracks in a music streaming platform
"!: playlist given in a query
ℳ""!: #music tracks from ℳ∖"!
%('): function to measure a fitness between 'and "!
<latexit sha1_base64="MHcNzBL3h5x6kzHiaSn8PUyr34Q=">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</latexit>
argmax
Mkpq
X
m2Mkpq
f(m)
Available information
nPlaylistsin a music streaming platform
nMusic tracks in a music streaming platform
nPropertiesof music tracks

7. 2. Related Work
7
1.Content-based
2.Collaborative Filtering

8. 8
Content-based method
Core: Measure similarity between tracks based on properties
likes
similar
probably likes
User
Property example: Artist, Genre, Released year

9. 9
Content-based method
Problem: Only partially capture playlist features
!A Whole New World
!Beauty and the Beast
!How Far I’ll Go
…
PlaylistMusical Tracks
Disney Movies
Popular Tracks
○well captured!
△partially captured!
×cannot captured!
Recommendation Result: Unpopular, non-Disney musical tracks

10. 10
Collaborative Filtering
Core: Measure similarity between playlists
contains
likely to contain
Playlist A
Playlist B
contains
contains

11. 11
Collaborative Filtering
Problem: It cannot consider properties, cold start
Recommendation System
!×50
Playlist
Well captured
recommendation
Results
!×3
Playlist
Low-quality
recommendation
Results
For playlists with a small number of tracks, the method cannot
measure similarity properly.

12. 3. Proposed Method
12
1.Tripartite Graph Construction
2.SoI Extraction
3.Biased PPR

13. 13
Overview
Idea: Consider both propertiesandplaylist similarities
1.Model the relationship based
on the graph structure
2. Extract SoI related to the
playlist given in the query3. Calculate PPR on SoI,
recommend !tracks

14. 14
Step 1. Tripartite Graph Construction
Playlists, Tracks, Properties of tracksas nodes
Playlist APlaylist BPlaylist C
HelloSee you againSugarPayphoneAnimals
AdeleCharlie PuthMaroon 5Pop (genre)Soul (genre)
Playlists
Tracks
Properties
This can consider both properties of tracks andplaylist similarities!

15. 15
Step 2. SoI Extraction
Subgraph of Interest (SoI):
nA subgraph on a tripartite graph
nContains nodes and edges from tracks in !)to 2-hop ahead
Playlists
Tracks
Properties

16. 16
Step 2. SoI Extraction
Extract tracksin !!
Playlists
Tracks
Properties

17. Playlists
Tracks
Properties
17
Step 2. SoI Extraction
Extract playlistsand propertiesconnected to tracksin !!

18. Playlists
Tracks
Properties
18
Step 2. SoI Extraction
Extract tracksconnected to playlists and properties
SoI excludes most of the graph not related to a given playlist!

19. 19
Step 3. Biased PPR
Personalized PageRank (PPR)
nScoring the influenceof each node on the interesting nodes
nWeight on interesting nodes àSpecified by Personalized Vector
012
43
GraphPersonalized Vector
(interested in node 0)
1
0
0
0
0
PPR0:0.308…
1:0.261…
2:0.074…
3:0.198…
4:0.158…
PPR value for each node

20. 20
Step 3. Biased PPR
Procedure
1.Specify equal weights for tracks in ."by personalized vector
2.Calculate PPR on the SoI, recommend top /tracks in order of PPR value
Playlists
Tracks
Properties
0.50.5

21. 21
Problems and Solutions in Simple PPR (1/2)
Problems:
Recommendations are biased toward unpopular tracks.
Solutions:
Give weights according to the degreeof nodes.
Popular tracks (highdegree):
PPR values of surrounding nodes are low
Cause of this problem
popularunpopular
Unpopular tracks (lowdegree):
PPR values of surrounding nodes are high

22. 22
Problems and Solutions in Simple PPR (2/2)
Problems:
User’s preferences for properties are not well reflected
Solutions:
Give weights to propertiesas well as tracks
Before: Weight on tracksonlyAfter: Weight on tracksand properties

23. 4. Experiments
23

24. 24
Dataset
Playlists:
n10,000playlists published in Spotify Million Playlist Dataset
Challenge*1
nSplit playlists in two parts:
pPlaylist group A: serves as a playlists on a music streaming platform
pPlaylist group B: serves as playlists given in a query (.")
nUsing the following properties:
pArtist, Album, Genre, Released year
https://www.aicrowd.com/challenges/spotify-million-playlist-dataset-challenge
TracksPlaylistsArtistsAlbumsGenres
170,08910,00035,89281,5523,865
Dataset statistics

25. 25
Experimental Settings
We focus on the accuracyof recommendations.
nHow? àPrepare ground truth for the recommendation results
Playlist Group B
Playlist (Query)
Input
Randomly select 10 tracks
from each playlist
recommend !tracks
Results
Ground truth
tracks not selected
evaluation
Playlist Group A
TracksProperties
Recommendation System

26. 26
Methods
1.HIN-MRS*1
pThe state-of-the-art content-based APC method
pExplores tracks similar to ."from the processed music graph by using
PPR.
2.Collaborative Filtering
pFilters music tracks similar to ."from a music-playlist bipartite graph.
3.Content-based method
pBaseline method
pMeasures similarity between music tracks and ."by counting up the
number of shared properties.
*1: Wang, R et al., Heterogeneous Information Network-Based Music Recommendation System in
Mobile Networks, Computer Communications, Vol. 150, pp. 429–437 (2020).

27. 27
Evaluation Criteria
1.Mean Average Precision (MAP)
pTaken from 0 to 1
pApproaching 1 if the recommendation results close to the ground truth.
2.Normalized Discounted Cumulated Gain (NDCG)
pTaken from 0 to 1
pApproaching 1 if the recommendation rank is appropriate.

28. 28
Recommendation accuracy by varying !
Thetripartite graph handles the relationship between
tracks, playlists, and properties of tracks!
MAP by varying # NDCG by varying #

29. 29
Recommendation accuracy by varying |#!|
The proposed method is more effective with a sufficient number of tracks.
Methods|$!|=5101520
Proposed method0.1920.2460.2600.258
HIN-MRS0.1960.2310.2380.243
Collaborative Filtering0.1520.1660.1660.164
Content-based0.0990.1220.1290.136
Methods|$!|=5101520
Proposed method0.6000.6610.6670.674
HIN-MRS0.6160.6480.6480.656
Collaborative Filtering0.5490.5700.5740.576
content-based0.4830.5110.5110.526
MAP, NDCG by varying |$!|
Proposed method causes cold startàFuture work

30. 30
Effectiveness of Biased PPR
Avoid the concentration of unpopular tracks at the
top of the recommendation results.
MAP NDCG

31. 31
ConclusionConclusion:
nWe proposed a novel music recommendation algorithm for APC.
nOur algorithm constructs a tripartite graph from tracks, playlists, and
properties.
nSoI and Biased PPR highlighted important nodes and improved
recommendation accuracy.
nOur experimental analysis demonstrated that our algorithm outperforms
the state-of-the-art methods regarding recommendation accuracy.
Future Work:
nImproving cold start
nLeveraging playlist names and user metadata

32. 32
Thank you!