33
The Benefit of Using Tag-Based Profiles LA-Web 2007 2008/03/14 Claudiu S. Firan Wolfgang Nejdl Raluca Paiu L3S Research Center

The Benefit of Using Tag-Based Profiles LA-Web 2007

  • Upload
    sherry

  • View
    26

  • Download
    0

Embed Size (px)

DESCRIPTION

The Benefit of Using Tag-Based Profiles LA-Web 2007. Claudiu S. FiranWolfgang NejdlRaluca Paiu L3S Research Center. 2008/03/14. Introduction. Collaborative tagging has emerged as an important way to organize, provide and share information about the resources on the Web. - PowerPoint PPT Presentation

Citation preview

Page 1: The Benefit of Using Tag-Based Profiles LA-Web 2007

The Benefit of UsingTag-Based Profiles LA-Web 2007

2008/03/14

Claudiu S. Firan Wolfgang Nejdl Raluca PaiuL3S Research Center

Page 2: The Benefit of Using Tag-Based Profiles LA-Web 2007

Introduction

Collaborative tagging has emerged as an important way to organize, provide and share information about the resources on the Web.

Recent research has shown that such tag distributions stabilize over time, and can be used to improve search on the Web.

How can they be used to enable personalized recommendations? More specifically : Music recommendations.

Page 3: The Benefit of Using Tag-Based Profiles LA-Web 2007

Current System

Collaborative Filtering Cold start problem Poor variety

Content Similarity Similarity does not imply preference

Hybrid Methods Complex for general users

Tags not used for recommendation

Page 4: The Benefit of Using Tag-Based Profiles LA-Web 2007

Related Work

Most music recommender systems are based on collaborative filtering

Other approaches exist FOAF : friend-of-a-friend, RSS { users, ratings, contents } Bayesian network But,

Profiles not automatically inferred from music data Profiles are track based, not tag based

Similar approach Bookmarking recommendation on Del.icio.us

Page 5: The Benefit of Using Tag-Based Profiles LA-Web 2007

Functionalities & Usage Data ( Last.fm )

Track User Tag

Page 6: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Track name Artist name Album name Tags & score Number of times has been played User comments

Page 7: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Page 8: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Page 9: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Page 10: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Page 11: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Page 12: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Page 13: The Benefit of Using Tag-Based Profiles LA-Web 2007

Track Data ( Last.fm )

Total 317,058 tracks

Page 14: The Benefit of Using Tag-Based Profiles LA-Web 2007

Tag Data ( Last.fm )

Number of times has been used Number of users have used Similar tags with scores

Page 15: The Benefit of Using Tag-Based Profiles LA-Web 2007

Tag Data ( Last.fm )

Page 16: The Benefit of Using Tag-Based Profiles LA-Web 2007

Tag Data ( Last.fm )

Total 21,177 tags

Page 17: The Benefit of Using Tag-Based Profiles LA-Web 2007

Tag Data ( Last.fm )

Page 18: The Benefit of Using Tag-Based Profiles LA-Web 2007

User Data ( Last.fm )

ID Gender Age Location Register date Number of tracks Friends, Neighbors, Groups Tags

Page 19: The Benefit of Using Tag-Based Profiles LA-Web 2007

User Data ( Last.fm )

Page 20: The Benefit of Using Tag-Based Profiles LA-Web 2007

User Data ( Last.fm )

Total 289,654 users Filter :

> 50 tracks > 10 tags

12,193 users left

Page 21: The Benefit of Using Tag-Based Profiles LA-Web 2007

User Profiles

Track-based A list of < track, score > pairs.

Tag-based A list of < tag, score > pairs.

Page 22: The Benefit of Using Tag-Based Profiles LA-Web 2007

User Profiles

Track-based Track-Tag-based Tag-based

Last.fm

user

Tracks that user played, and times

as score.

•Track list•Tags by all user•# times played (user) # times tagged (all)

Tags used by user# times tagged

(user)

Non-Last.f

muser

Tracks from user’s PC, and times played by all

Last.fm users.

•Track list•Tags by all user•# times played (all) # times tagged (all)

# tracks on user’s PC

( top 29 )

Page 23: The Benefit of Using Tag-Based Profiles LA-Web 2007

Music Recommendations

Collaborative Filtering based on Tracks baseline

Collaborative Filtering based on Tags CFTTI, CFTTN, CFTG

Search based on Tags STTI, STTN, STG

Page 24: The Benefit of Using Tag-Based Profiles LA-Web 2007

Music Recommendations

Track-based Track-Tag-based Tag-based

Collaborative Filtering

CFTR10 similar usercompute tracksrecommend

CFTTICFTTN

CFTG

Search

STTISTTN

STG

Page 25: The Benefit of Using Tag-Based Profiles LA-Web 2007

Music Recommendations

user

all user

similar

user

recommend track

Lucene similarit

y

tracktag

Page 26: The Benefit of Using Tag-Based Profiles LA-Web 2007

Music Recommendations

userR

all user

similar

user

Rrecommend track

Lucene similarit

y

tracktag

CFTR

Page 27: The Benefit of Using Tag-Based Profiles LA-Web 2007

Music Recommendations

userR

G

all userR

G

similar

userG

recommend track

Lucene similarit

y

R

tracktagCFTTI, CFTTN

CFTG

Page 28: The Benefit of Using Tag-Based Profiles LA-Web 2007

Music Recommendations

userR

G

all userR

G

recommend track

Lucene similarit

y

R

tracktag

STG

STTI, STTN

Page 29: The Benefit of Using Tag-Based Profiles LA-Web 2007

Evaluation

7 variant algorithms each returns 10 recommends

18 subject to rate Rating

Preference : 0, 1, 2 Novelty : 0, 1, 2

NDCG, Popularity, Novelty

Page 30: The Benefit of Using Tag-Based Profiles LA-Web 2007

Results - 1

Page 31: The Benefit of Using Tag-Based Profiles LA-Web 2007

Results - 2

Page 32: The Benefit of Using Tag-Based Profiles LA-Web 2007

Results - 3

Page 33: The Benefit of Using Tag-Based Profiles LA-Web 2007

Conclusions

Analyze tag usage of the most popular music community site, Last.fm

Compare user profiles based on tags with conventional based on tracks

Specify recommendation algorithms based on tag-based user profiles