Query-based Music Recommendations via Preference Embedding - - PowerPoint PPT Presentation

▶
query based music recommendations via preference embedding
SMART_READER_LITE
LIVE PREVIEW

Query-based Music Recommendations via Preference Embedding - - PowerPoint PPT Presentation

Jan 10, 2023 4.02k likes •4.36k views

Query-based Music Recommendations via Preference Embedding Chih-Ming Chen, Ming-Feng Tsai, Yu-Ching Lin, Yi-Hsuan Yang. Institutes involved in this research work CLIP Lab, National Chengchi University MAC Lab,

slide-1
SLIDE 1

Query-based Music Recommendations via Preference Embedding

Chih-­‑Ming ¡Chen, ¡Ming-­‑Feng ¡Tsai, ¡Yu-­‑Ching ¡Lin, ¡Yi-­‑Hsuan ¡Yang.

slide-2
SLIDE 2

Institutes involved in this research work

MAC Lab, CITI, Academia Sinica CLIP Lab, National Chengchi University Machine Learning Team, KKBOX

slide-3
SLIDE 3

Query-based Music Recommendations

slide-4
SLIDE 4

Query-based Music Recommendations

slide-5
SLIDE 5

Query-based Music Recommendations

slide-6
SLIDE 6

Query-based Music Recommendations

slide-7
SLIDE 7

Query-based Music Recommendations

slide-8
SLIDE 8

Query-based Music Recommendations

slide-9
SLIDE 9

Latent Space

The Graph Embedding Models

Vertices Relation Graph

slide-10
SLIDE 10

Vertices

How to build
 the Relation Graph? How to learn
 the representations?

The Graph Embedding Models

slide-11
SLIDE 11

Construction of User Preference Network

User Track Album Artist

U1 T1 T2 T3 T4 T5 U2 U3 T6 U4

112 16 8

  • 2

119 64

  • 32
  • 109

5

  • 12

8

# of Listening / Rating /
 Like / Dislike / …

slide-12
SLIDE 12

User Track Album Artist

U1 T1 T2 T3 T4 T5 U2 U3 Al1 Al2 Al3 Al4 T6 U4

112 16 8

  • 2

119 64

  • 32
  • 109

5

  • 12

8 112 24

  • 121 64
  • 32
  • 114
  • 20

Construction of User Preference Network

slide-13
SLIDE 13

User Track Album Artist

U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

112 16 8

  • 2

119 64

  • 32
  • 109

5

  • 12

8 112 24

  • 121 64
  • 32
  • 114
  • 20

112 24

  • 121

64

  • 32

114

  • 20

Construction of User Preference Network

slide-14
SLIDE 14

User Preference Network

U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

Edges: User Preference Bipartite Graph Heterogeneous Graph

it’s similar to CF-based models binary value / numerical value it considers multiple entities

slide-15
SLIDE 15

User Preference Network

U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

Edges: User Preference Bipartite Graph Heterogeneous Graph

it’s similar to CF-based models binary value / numerical value it considers multiple entities

This is how we achieve the
 Query-based recommendations

slide-16
SLIDE 16

Heterogeneous Preference Embedding (HPE)

U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

Pr( community( ) | )

U3

Φ( )

compress the info

U3

slide-17
SLIDE 17

Heterogeneous Preference Embedding (HPE)

U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

Pr( community( ) | )

U3

Φ( )

U3

Pr( | )

U3 T3

Sample an Edge

slide-18
SLIDE 18

Heterogeneous Preference Embedding (HPE)

Pr( | )

U3 T3

Sample an Edge

Pr( community( ) | )

U3

Φ( )

U3

Random Walk

Pr( | )

U3 U1

Pr( | )

U3 Al1 U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

slide-19
SLIDE 19

Heterogeneous Preference Embedding (HPE)

Pr( | )

U3 T3

Sample an Edge

Pr( community( ) | )

U3

Φ( )

U3

Random Walk

Pr( | )

U3 U1

Pr( | )

U3 Al1 U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

O = X

(i,j)2S

wi,j log p(vj|Φ(vi)) + λ X

i

kΦ(vi)k2

+ negative sampling

slide-20
SLIDE 20

Heterogeneous Preference Embedding (HPE)

Pr( | )

U3 T3

Sample an Edge

Pr( community( ) | )

U3

Φ( )

U3

Random Walk

Pr( | )

U3 U1

Pr( | )

U3 Al1 U1 T1 T2 T3 T4 T5 Ar1 U2 U3 Ar2 Ar3 Al1 Al2 Al3 Al4 T6 U4

O = X

(i,j)2S

wi,j log p(vj|Φ(vi)) + λ X

i

kΦ(vi)k2

+ negative sampling

slide-21
SLIDE 21

Performance of Preference Embedding

HitRatio@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 2.66% 2.66% 2.66% 4.32% 4.32% 4.32% 0.92% 0.92% 0.92% MF 3.02% 3.93% 4.22% 7.11% 8.49% 8.93% 1.37% 1.79% 2.00% DeepWalk 3.18% 3.55% 3.54% 11.61% 12.55% 13.08% 1.71% 1.95% 1.95% LINE-2nd 3.44% 3.74% 4.10% 12.79% 13.47% 12.77% 1.62% 1.60% 1.14% Proposed PE 3.54% *4.22% 4.51% 12.95% *13.74% *14.20% *2.08% *2.15% *2.19% mAP@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 3.27% 3.27% 3.27% 5.03% 5.03% 5.03% 1.04% 1.04% 1.04% MF 1.87% 2.34% 2.60% 4.65% 5.85% 6.16% 1.88% 2.44% 2.81% DeepWalk 1.82% 2.10% 1.99% 8.73% 9.47% 10.01% 2.66% 2.70% 2.55% LINE-2nd 2.00% 2.10% 2.38% 9.95% 10.64% 10.09% 1.84% 1.60% 1.44% Proposed PE 2.08% 2.55% 2.71% 10.14% 10.86% *11.31% 2.86% *3.09% *3.12%

slide-22
SLIDE 22

Performance of Preference Embedding

HitRatio@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 2.66% 2.66% 2.66% 4.32% 4.32% 4.32% 0.92% 0.92% 0.92% MF 3.02% 3.93% 4.22% 7.11% 8.49% 8.93% 1.37% 1.79% 2.00% DeepWalk 3.18% 3.55% 3.54% 11.61% 12.55% 13.08% 1.71% 1.95% 1.95% LINE-2nd 3.44% 3.74% 4.10% 12.79% 13.47% 12.77% 1.62% 1.60% 1.14% Proposed PE 3.54% *4.22% 4.51% 12.95% *13.74% *14.20% *2.08% *2.15% *2.19% mAP@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 3.27% 3.27% 3.27% 5.03% 5.03% 5.03% 1.04% 1.04% 1.04% MF 1.87% 2.34% 2.60% 4.65% 5.85% 6.16% 1.88% 2.44% 2.81% DeepWalk 1.82% 2.10% 1.99% 8.73% 9.47% 10.01% 2.66% 2.70% 2.55% LINE-2nd 2.00% 2.10% 2.38% 9.95% 10.64% 10.09% 1.84% 1.60% 1.44% Proposed PE 2.08% 2.55% 2.71% 10.14% 10.86% *11.31% 2.86% *3.09% *3.12%

slide-23
SLIDE 23

Performance of Preference Embedding

HitRatio@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 2.66% 2.66% 2.66% 4.32% 4.32% 4.32% 0.92% 0.92% 0.92% MF 3.02% 3.93% 4.22% 7.11% 8.49% 8.93% 1.37% 1.79% 2.00% DeepWalk 3.18% 3.55% 3.54% 11.61% 12.55% 13.08% 1.71% 1.95% 1.95% LINE-2nd 3.44% 3.74% 4.10% 12.79% 13.47% 12.77% 1.62% 1.60% 1.14% Proposed PE 3.54% *4.22% 4.51% 12.95% *13.74% *14.20% *2.08% *2.15% *2.19% mAP@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 3.27% 3.27% 3.27% 5.03% 5.03% 5.03% 1.04% 1.04% 1.04% MF 1.87% 2.34% 2.60% 4.65% 5.85% 6.16% 1.88% 2.44% 2.81% DeepWalk 1.82% 2.10% 1.99% 8.73% 9.47% 10.01% 2.66% 2.70% 2.55% LINE-2nd 2.00% 2.10% 2.38% 9.95% 10.64% 10.09% 1.84% 1.60% 1.44% Proposed PE 2.08% 2.55% 2.71% 10.14% 10.86% *11.31% 2.86% *3.09% *3.12%

slide-24
SLIDE 24

Performance of Preference Embedding

HitRatio@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 2.66% 2.66% 2.66% 4.32% 4.32% 4.32% 0.92% 0.92% 0.92% MF 3.02% 3.93% 4.22% 7.11% 8.49% 8.93% 1.37% 1.79% 2.00% DeepWalk 3.18% 3.55% 3.54% 11.61% 12.55% 13.08% 1.71% 1.95% 1.95% LINE-2nd 3.44% 3.74% 4.10% 12.79% 13.47% 12.77% 1.62% 1.60% 1.14% Proposed PE 3.54% *4.22% 4.51% 12.95% *13.74% *14.20% *2.08% *2.15% *2.19% mAP@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 3.27% 3.27% 3.27% 5.03% 5.03% 5.03% 1.04% 1.04% 1.04% MF 1.87% 2.34% 2.60% 4.65% 5.85% 6.16% 1.88% 2.44% 2.81% DeepWalk 1.82% 2.10% 1.99% 8.73% 9.47% 10.01% 2.66% 2.70% 2.55% LINE-2nd 2.00% 2.10% 2.38% 9.95% 10.64% 10.09% 1.84% 1.60% 1.44% Proposed PE 2.08% 2.55% 2.71% 10.14% 10.86% *11.31% 2.86% *3.09% *3.12%

slide-25
SLIDE 25

Performance of Preference Embedding

HitRatio@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 2.66% 2.66% 2.66% 4.32% 4.32% 4.32% 0.92% 0.92% 0.92% MF 3.02% 3.93% 4.22% 7.11% 8.49% 8.93% 1.37% 1.79% 2.00% DeepWalk 3.18% 3.55% 3.54% 11.61% 12.55% 13.08% 1.71% 1.95% 1.95% LINE-2nd 3.44% 3.74% 4.10% 12.79% 13.47% 12.77% 1.62% 1.60% 1.14% Proposed PE 3.54% *4.22% 4.51% 12.95% *13.74% *14.20% *2.08% *2.15% *2.19% mAP@10 lastfm-1k (window=5) KKBOX (window=5) MSD (original) d = 16 d = 32 d = 64 d = 64 d = 128 d = 256 d = 64 d = 128 d = 256 Popularity 3.27% 3.27% 3.27% 5.03% 5.03% 5.03% 1.04% 1.04% 1.04% MF 1.87% 2.34% 2.60% 4.65% 5.85% 6.16% 1.88% 2.44% 2.81% DeepWalk 1.82% 2.10% 1.99% 8.73% 9.47% 10.01% 2.66% 2.70% 2.55% LINE-2nd 2.00% 2.10% 2.38% 9.95% 10.64% 10.09% 1.84% 1.60% 1.44% Proposed PE 2.08% 2.55% 2.71% 10.14% 10.86% *11.31% 2.86% *3.09% *3.12%

slide-26
SLIDE 26

Extension Work

slide-27
SLIDE 27

Extension Work

Multiple
 Queries:

slide-28
SLIDE 28

Extension Work

Multiple
 Queries: