This presentation is about our paper which was presented at the Hypertext conference 2012. In this paper, we investigate a methodology for measuring the influence of tag recommenders on the indexing quality in collaborative tagging systems. We propose to use the inter-resource consistency as an indicator of indexing quality. The inter-resource consistency measures the degree to which the tag vectors of indexed resources reflect how the users understand the resources. We use this methodology for evaluating how tag recommendations coming from (1) the popular tags at a resource or from (2) the user's own vocabulary influence the indexing quality. We show that recommending popular tags decreases the indexing quality and that recommending the user's own vocabulary increases the indexing quality.
Links to the paper:
http://dx.doi.org/10.1145/2309996.2310009
http://www.west.uni-koblenz.de/files/publications/dellschaft2012mti.pdf
Measuring the Influence of Tag Recommenders on the Indexing Quality in Tagging Systems
1. Web Science & Technologies
University of Koblenz ▪ Landau, Germany
Measuring the Influence of Tag Recommenders
on the Indexing Quality in Tagging Systems
Klaas Dellschaft
klaasd@uni-koblenz.de
Steffen Staab
staab@uni-koblenz.de
2. Collaborative Tagging Systems
Objectives of tag recommenders:
Improve indexing quality ⇒ retrieval results
Reduce tagging effort
Measuring the Influence of Tag Recommenders Slide 2 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
3. Outline
Measures of indexing quality
What to understand under “indexing quality”?
Inter-resource consistency ⇔ inter-indexer consistency
Evaluation of the measures
Are the measures correlated with each other?
User study: Apply measures for two recommenders
Evaluation results
Conclusions
Measuring the Influence of Tag Recommenders Slide 3 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
4. Measures of
Indexing Quality
Measuring the Influence of Tag Recommenders Slide 4 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
6. Measures of indexing quality
Inter-resource consistency
Compare resource similarity to the tag vector distance
Requires external knowledge about similarity of resources
Direct but sophisticated measure of indexing quality
Inter-indexer consistency
Do users agree on common description for a resource?
Assumption: Users select tags independent of each other
Indirect but easy measure of indexing quality
Which measure to use for evaluating tag recommenders?
Measuring the Influence of Tag Recommenders Slide 6 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
7. Research Hypotheses
Hypothesis: Inter-indexer consistency does not measure the
influence of tag recommenders on the indexing quality!
Popular Tags: Suggest most popular tags of a resource
H1a: Popular Tags increase the inter-indexer consistency
H1b: Popular Tags decrease the inter-resource consistency
User Tags: Suggest all tags previously applied by the user
H2a: User Tags lead to a decreased or unchanged inter-indexer
consistency
H2b: User Tags increase the inter-resource consistency
The measures do not correlate when evaluating tag recommenders
Measuring the Influence of Tag Recommenders Slide 7 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
8. Measuring Inter-Resource Consistency
Idea: Compare resource similarity and tag vector distance
ai: Average distance to resources in the same cluster
bi: Average distance to resources in the closest other cluster
bi − ai
si =
max(ai , bi )
resource
cluster of similar resources
inconsistent consistent even more
-1 0 consistent +1
Measuring the Influence of Tag Recommenders Slide 8 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
9. Measuring Inter-Indexer Consistency
Idea: Do users agree on common description for a resource?
Tag Reuse Rate
Average number of users who apply a tag
Used in the related work
news 8 8 8
humor 4 6 6
fun 2 2 0
patents 0 0 0
Tag Reuse Rate: 4.7 5.3 7
Measuring the Influence of Tag Recommenders Slide 9 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
10. Evaluation
Measuring the Influence of Tag Recommenders Slide 10 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
11. Experimental Setup
Objective:
Are inter-resource and inter-indexer correlated if tag
recommendations are given?
Task given to users:
Assign keywords to 10 web pages.
After tagging, cluster web pages according to their
similarity (⇒ inter-resource consistency).
Three different experimental conditions:
No Suggestions
User Tags
Popular Tags
Further divided into an English and German user group
Measuring the Influence of Tag Recommenders Slide 11 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
12. Suggestion of Popular Tags – Screenshot
Measuring the Influence of Tag Recommenders Slide 12 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
13. Clustering of Similar Web Pages – Screenshot
Measuring the Influence of Tag Recommenders Slide 13 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
14. Results
Measuring the Influence of Tag Recommenders Slide 14 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
15. Sizes of the Tagging Data Set
German User Group:
#Users #Tags #TAS #TAS / #User
No Suggestions 74 706 2134 28.84
Popular Tags 78 531 2228 28.56
User Tags 79 466 1507 19.08
English User Group:
#Users #Tags #TAS #TAS / #User
No Suggestions 115 973 3150 27.39
Popular Tags 118 550 3003 25.45
User Tags 118 819 2919 24.74
Measuring the Influence of Tag Recommenders Slide 15 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
16. The Clustering Data Set
In average, each user identified 4.59 clusters
Overall, 146 distinct clusters have been identified
11 most frequent clusters ⇒ 70% of the data
The web pages cover ~7 topics
3 web pages are on the border between two topics
Measuring the Influence of Tag Recommenders Slide 16 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
17. Differences in the Topical Clusters
Cluster probabilities in English experiment
No Suggestions
Popular Tags
User Tags
The Onion + BBC The Onion + Patents
⇒ News ⇒ Humor
English Popular Tags condition has to be excluded
Measuring the Influence of Tag Recommenders Slide 17 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
18. Measuring the Inter-Resource Consistency
H1a: Popular Tags decrease the inter-resource consistency
H2a: User Tags increase the inter-resource consistency
Expectation: E(spt,i) < E(sns,i) < E(sut,i)
E(spt,i) E(sns,i) E(sut,i)
German Users 0.1474 0.1847 0.2367
English Users N/A 0.1713 0.1915
(All differences are significant!)
Measuring the Influence of Tag Recommenders Slide 18 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
19. Measuring the Inter-Indexer Consistency
H1b: Popular Tags increase the inter-indexer consistency
H2b: User Tags lead to a decreased or unchanged
inter-indexer consistency
Expectation: E(trpt,i) > E(trns,i) ≥ E(trut,i)
E(trpt,i) E(trns,i) E(trut,i)
German Users 3.60 2.44 2.39*
English Users 4.67 2.76 2.68*
* Differences between E(trns,i) and E(trut,i) not significant
Measuring the Influence of Tag Recommenders Slide 19 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
20. Conclusions
Measures of indexing quality
Inter-resource consistency
Inter-indexer consistency
Measures do not correlate if recommendations are given
Only inter-resource consistency can be used
Popular Tags
Do not lead to consistent descriptions across resources
Are rather counterproductive for indexing resources
User Tags
Lead to consistent descriptions across resource
Consolidate the personomy of users
Measuring the Influence of Tag Recommenders Slide 20 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de
21. Paper:
K. Dellschaft & S. Staab. Measuring the Influence of Tag
Recommenders on the Indexing Quality in Tagging Systems.
Proceedings of the Hypertext Conference, 2012
http://dl.acm.org/citation.cfm?id=2310009
Experimental Interface:
http://userpages.uni-koblenz.de/~klaasd/experiment/
Data Set:
http://west.uni-koblenz.de/Research/DataSets/tagging-experiment/
Measuring the Influence of Tag Recommenders Slide 21 of 21
Klaas Dellschaft (klaasd@uni-koblenz.de) http://west.uni-koblenz.de