The document describes Dedalo, a system that automatically explains clusters of data by traversing linked data to find explanations. It evaluates different heuristics for guiding the traversal, finding that entropy and conditional entropy outperform other measures by reducing redundancy and search time. Experiments on authorship clusters, publication clusters, and library book borrowings demonstrate Dedalo's ability to discover explanatory linked data patterns within a limited domain. Future work includes extending Dedalo to handle more complex datasets by addressing issues such as sameAs linking and use of literals.

1. Dedalo: looking for Clusters' Explanations in a
Labyrinth of Linked Data
Ilaria Tiddi, Mathieu d'Aquin, Enrico Motta
Knowledge Media Institute, The Open University
May 28, 2014

2. The Knowledge Discovery process
Explaining patterns requires background knowledge.
Background knowledge is attributed to the experts.
Background knowledge comes from dierent domains.
Experts might not be aware of some background knowledge.

3. Explaining clusters: an example
Authors clustered according to the papers they wrote together.
How to explain those clusters?

4. Explaining clusters { the easy solution
Use an expert

5. Explaining clusters { the easy solution
Use an expert
each cluster represents a research group in KMi

6. Explaining clusters { the easy solution
Use an expert
each cluster represents a research group in KMi
Can one trust those experts?

7. Explaining clusters { the nice solution
Use Inductive Logic Programming (ILP)
E+ (positive examples) E (negative examples)
attendsESWC(M.dAquin).
attendsESWC(E.Motta).
attendsESWC(V.Lopez).
B: knowledge about E = E+ [ E
submitted(M.dAquin). submitted(V.Lopez).
submitted(E.Motta).
accepted(V.Lopez). accepted(M.dAquin).
Learn a complete (B [ H E+) and consistent (B [ H 2 E)
explanation for the relation attendsESWC(X).
attendsESWC(X) - submitted(X)^accepted(X)

8. Explaining clusters { still the nice solution
E+ (positive examples) E (negative examples)
inMyCluster(M.dAquin).
inMyCluster(M.Fernandez).
inMyCluster(V.Lopez).
inMyCluster(H.Saif).
inMyCluster(M.Sabou).
inMyCluster(C.Pedrinaci).
inMyCluster(J.Domingue).
B: knowledge about E = E+ [ E
B?
inMyCluster(X) { ?

9. Explaining clusters { the cool solution
Integrate ILP with Linked Data

10. Explaining clusters { the cool solution
E+ (positive examples) E (negative examples)
inMyCluster(M.dAquin).
inMyCluster(M.Fernandez).
inMyCluster(V.Lopez).
inMyCluster(H.Saif).
inMyCluster(M.Sabou).
inMyCluster(C.Pedrinaci).
inMyCluster(J.Domingue).
B: knowledge about E = E+ [ E
topic(M.dAquin, SemanticWeb). topic(M.Sabou, SemanticWeb).
topic(V.Lopez, SemanticWeb). topic(H.Saif, SocialWeb).
topic(C.Pedrinaci, SemanticWebServices).
topic(J.Domingue, SemanticWebServices).
topic(M.Fernandez, SocialWeb).
inMyCluster(X) - topic(X,SemanticWeb)
Is this enough?

11. Producing Linked Data Explanations
on similar topics
People working in the same place are likely to write papers together.
on the same project

12. Producing Linked Data Explanations
People working
under the same person
are likely to write papers together.
with the same partner

13. Producing Linked Data Explanations
People working under people interested in the same thing write papers together.

14. Integrating ILP and Linked Data
Add to B each Linked Data explanation hi = hpk i:hvk i*,
where:
pk (path): a chain of RDF properties
pk = fprop0 ! prop1 ! : : : ! propng
vk (value): a

15. nal instance
roots(hi ): elements 2 Ci having hi in common
roots(hi)=fou:M.dAquin, ou:V.Lopez, ou:M.Saboug
*spread across dierent datasets
hi = hou:project!ou:ledBy!foaf:topicipk :hedu:SemanticWebivk
Building each hi :
{ how?
{ which chains of properties?
{ where to

16. nd the good ones?

17. Dedalo { An iterative Linked Data traversal
Scoring hypotheses
WRacc
1(hi ) = jroots(hi )j
jRj
jroots(hi )Ci j
jroots(hi )j jCi j
jRj
1 Geng et al. (2006). Interestingness measures for data mining: A survey.

18. Dedalo { An iterative Linked Data traversal
How to de

19. ne the interestingness of a path pk?
How to reach the best hi in the shortest time?

20. Dedalo { Comparing Heuristics
We chose to compare dierent strategies.
We want to

21. nd the path pk leading to the best hi in the shortest time.
We want to save time and computational complexity
Path Length length of pk in number of properties composing it
Path Frequency frequency of the paths in the graph
Adapted PMI joint and individual distribution of pk and Ci
Adapted TF{IDF how important is pk (term) in Ci (doc)
Delta jvals(pk )j jCj
Entropy2 distribution of jvals(pk )j
Conditional Entropy distribution of jvals(pk )j w.r.t. Ci
2Shannon, C. (1948). A Mathematical Theory of Communication.

22. Dedalo's Heuristics
Ci=fou:M.dAquin, ou:V.Lopez, ou:M.Saboug
Path Frequency top(pk)=hfoaf:topici

23. Dedalo's Heuristics
Ci=fou:M.dAquin, ou:V.Lopez, ou:M.Saboug
Adapted TF{IDF top(pk)=hou:exMember i

24. Dedalo's Heuristics
Ci=fou:M.dAquin, ou:V.Lopez, ou:M.Saboug
Entropy top(pk)=hou:project!ou:ledBy!foaf:topici

25. Experiments { KMi co-authorship
Authors clustered according to their co-authorships.
Network Partitioning clustering, jRj=92, jCj= 6
Cycles
Wracc
0 5 10 15
0.00 0.04 0.08 0.12
Semantic Web authors
Len
Fq
D
Ent
C.Ent
TFIDF
PMI
Cycles
Wracc
0 5 10 15
0.00 0.04 0.08 0.12
Learning Analytics authors
Len
Fq
D
Ent
C.Ent
TFIDF
PMI
jCi j hi WRacc
22 horg:hasMembership!ox:hasPrincipal-
0.128
Investigator!org:hasMembershipip.hou:SmartProductsiv1
23 horg:hasMembership!ox:hasPrincipalInvestigator
0.127
!org:hasMembershipip.hou:SocialLearniv2

26. Experiments { KMi Publications
Papers clustered according to their keywords.
XK-Means clustering, jRj=865, jCj= 6
Cycles
Wracc
0 2 4 6 8 10
0.00 0.01 0.02 0.03 0.04 0.05
Learning Analytics papers
Len
Fq
D
Ent
C.Ent
TFIDF
PMI
Cycles
Wracc
0 2 4 6 8 10
0.00 0.02 0.04 0.06 0.08 0.10
Semantic Web papers
Len
Fq
D
Ent
C.Ent
TFIDF
PMI
jCi j hi WRacc
601 hdc:creator!ntag:isRelatedToip.hou:LearningAnalyticsiv1 0.042
220 hdc:creator!ntag:isRelatedToip.hou:SemanticWebiv2 0.073

27. Experiments {Hudders

28. eld's dataset
Books clustered according to the students' Faculties.
K-Means clustering, jRj=6969, jCj= 14
Cycles
Wracc
0 5 10 15
0.000 0.001 0.002 0.003 0.004 0.005
Music students' borrowings
Len
Fq
D
Ent
C.Ent
TFIDF
PMI
Cycles
Wracc
0 5 10 15
0.000 0.005 0.010 0.015
Theatre students' borrowings
Len
Fq
D
Ent
C.Ent
TFIDF
PMI
jCi j hi WRacc
335 hdc:subject!skos:broaderip1 .hlcsh:PhysicalScienceiv 0.005
919 hdc:creator!bl:hasCreated!dc:subjectip2 .hbl:EnglishDramaiv 0.013

29. Experiments { Comparing heuristics
Heuristics speed comparison in seconds.
KMiA1 KMiA2 KMiP1 KMiP2 Hud1 Hud2
Len 1.64 4.15 8.95 9.01 69.13 135.5
Freq 2.57 4.35 7.5 9.29 180 180
PMI 2.05 3.88 11.28 18.42 180 180
TF{IDF 1.69 3.18 10.61 17.19 180 180
Delta 2.02 3.92 180 180 180 180
Entropy 4.19 3.27 7.1 7.3 41.15 105.09
Conditional Entropy 2.64 3.89 7.48 7.55 70.91 40.89
/ { Len, Freq : fast but inaccurate baselines
, { Entropy/Conditional Entropy: outperforming measures,
reducing redundancy (following wrong paths) and time eorts
/ { PMI , TFIDF, Delta : they might work on less homogeneous
clusters

30. Conclusions
Linked Data { automatically explaining clusters
Dedalo { traversing Linked Data to reveal explanations
Entropy { driving the search in the Linked Data cloud
Beyond Dedalo.
Dedalo works as far as there is a limited domain.
New use-cases require its extension.

31. Future work: the OU students enrolment dataset
Add sameAs linking
Use of literals
Aggregation of atomic rules
Explore new hypotheses evaluation measures

32. Thanks for your attention!3
ilaria.tiddi@open.ac.uk
mathieu.daquin@open.ac.uk
enrico.motta@open.ac.uk
Questions?
Better asking the robot than the experts
3Special thanks to the KMi (happy) faces.