Title: Personalized%20Folksonomies%20Based%20on%20Hierarchical%20Tag%20Clustering
1Personalized Folksonomies Based on Hierarchical
Tag Clustering
A. Shepitsen, J. Gemmell, B. Mobasher and R.
Burke
2Agenda
- Overview of collaborative tagging systems
- Search and navigation in Folksonomies
- Personalized navigation in Folksonomies
- Hierarchical agglomerative tag clustering
- Experimental results
- Conclusions
3Storing the Resources Locally
Resource
Tag
4Tagging the Resource on Social Tagging System
Resource
Tag
5Delicious User Profile
User
Resources
Tags
6Last.Fm User Profile
7Navigation in Folksonomies
8Search in Folksonomies
Search Tag
Resources
9Tag Redundancy
C_sharp
C
Italian_Food
Italian_cuisine
DePaul_University
DePaul
10Tag Ambiguity
Java
Run
11Advantages of Personalization Based on Clustering
- Electiveness in treating tag redundancy
Eclipse
Sun
Java/JSEE
_Java
Java5
Java
_Java
Enigm
C
Java
Java/JSEE
String
JSP
Java5
12Advantages of PersonalizationBased on Clustering
- Electiveness in treating Tag ambiguity
Ambiguous Tag Apple (Flikr social tagging system)
13Personalization with clusters
Java
Coffee Drinker
Java
Coffee
StarBucks
Nestle
Nascafe
Morning_drink
Programmer
Java
Eclips
C
JSEE
C
Traveler
Java
West_Malaizija
Rest_Bruney
Indonesia_tours
Australia
14Tag Similarities Measurement
R1 R2 R3
Rn
Tag1 Tag2 Tag3 Tagn
5 7 14 15
11 6 12 8
7 9 7 14
8 12 5 10
15Similarity in non-descriptive tags/IDF
Log(N/n)
Cool
R1 R2
Rn
67 27 119
3.11 1.12 3.67
Log(N/n)
Antropology
R1 R2
Rn
29 12 23
4.35 2.7 3.78
16Agglomerative Example (Step and matrix)
Search Tag Design
.25
.4
.55
.7
.85
1.0
Step .15
Java
java
j2ee
Web
Deals
Tools
Search
Google
Design
SearchEng
webDesign
Shopping
Coffee
Bargins
Espresso
Food
Programming
17Cluster Cohesiveness
Single Linkage
Cluster B
Cluster A
Maximal Linkage
Cluster B
Cluster A
18Cluster Cohesiveness
Average Centroid Linkage
Cluster B
Cluster A
19Generalization Coefficient
Search Tag Java Generalization coefficient
2
.25
.4
.55
.7
.85
1.0
Step .15
java
Java
j2ee
Web
Deals
Tools
Search
Google
Design
SearchEng
webDesign
Shopping
Coffee
Bargins
Espresso
Food
Programming
20Division Coefficient
Division Coefficient0.6
.25
.4
.55
.7
.85
1.0
Step .15
java
Java
j2ee
Web
Deals
Tools
Search
Design
Google
SearchEng
webDesign
Shopping
Coffee
Bargins
Espresso
Food
Programming
Cluster1
Cluster2
Cluster3
Cluster4
Cluster5
Cluster6
21Personalization
Java
Coffee
Programmimg
1
StarBucks
Nestle
URL
Nascafe
Morning_drink
Programmimg
2
URL
Java
Coffee
Eclips
C
3
URL
JSEE
C
Tourism
4
URL
Java
West_Malaizija
Rest_Bruney
Coffee
5
URL
Indonesia_tours
Australia
22Leave one out approach
Query Tag
User
Tag Resourse
T1 T2 T3 T4 T5
R1 R2 R3 R4 R5
Target Resource
23Personalization Explanation
Coffee Drinker
Java
Coffee
StarBucks
Nestle
Morning_drink
Nascafe
Programmer
Java
Eclips
C
C
JSEE
Java
Traveler
Java
Rest_Bruney
West_Malaizija
Indonesia_tours
Australia
24Delicious Dataset
25Step Coefficient chart
26Generalization Coefficient Chart
27Division coefficient chart
28Maximal Complete Link
Cluster1
Tools
Google Tools SearchEng
Google
SearchEng
Espresso
Web
Food
Cluster2
Java
Tools Web Java
Cluster3
Espresso Food
29Maximal Complete Link Clustering
30K-Means ClusteringK
31Comparison of Clustering techniques
32Conclusions Future Work
Conclusions
- Clustering is an effective means for overcoming
tag ambiguity and tag redundancy - Hierarchical agglomerative clustering is found to
be the most effective clustering technique - Clustering can be used effectively for other
purposes in Folksonomies such as recommending
tags, resources and users
Future Work
- using PLSA and PCA to find the connection
between users and resources -
- using clusters for recommendation purposes
- implementing the notion of authority of users,
tags - and resources in Folksonomies
33Q/A?