ruleml2015 - using substitutive itemset mining framework for finding synonymous properties in linked...

Using Substitutive Itemset Mining Framework forFinding Synonymous Properties in Linked Data

Miko laj Morzy, Agnieszka Lawrynowicz, Mateusz Zozulinski

Poznan University of Technology, Poland

August 3rd, 2015RuleML 2015

Miko laj Morzy, Agnieszka Lawrynowicz, Mateusz Zozulinski ( Poznan University of Technology, Poland )Using Substitutive Itemset Mining Framework for Finding Synonymous Properties in Linked DataAugust 3rd, 2015 RuleML 2015 1

/ 15

Outline

Motivating Scenario

Substitutive Sets Mining

Finding Synonymous Properties with Substitutive Sets Mining

Summary


/ 15

Motivating scenario 1/3

Wiki (mappings

Wikipedia infoboxes -‐> DBpedia ontology)

Norah_Jones

Denton,_Texas dbpedia-‐prop:origin

Mark_Knopfler Gosforth dbpedia-‐owl:hometown

Peter_Gabriel Godalming dbpedia-‐owl:hometown


/ 15


DBpedia 2014 ontology has 1310 object and 1725 data properties

Many large Linked Data use relatively lightweight schemas with ahigh number of object properties


/ 15


Wiki (mappings

Wikipedia infoboxes -‐> DBpedia ontology)

dbpedia-‐owl:MusicalArAst dbpedia-‐owl:PopulatedPlace

Norah_Jones

Denton,_Texas dbpedia-‐prop:origin

Mark_Knopfler Gosforth dbpedia-‐owl:hometown

Peter_Gabriel Godalming dbpedia-‐owl:hometown




/ 15

Substitutive Sets Mining Framework

Frequent(Itemset(Mining(

Subs1tu1ve(Set(Genera1on(

Transac1on(DB(


/ 15

Frequent Itemsets

I = {i1, i2, . . . , im} - a set of items

DT = {t1, t2, . . . , tn}, where ∀iti ⊆ I -a database of transactions

support(X ) = ∣{t∈DT ∶X⊆t}∣∣DT ∣

ID Items

1 Nachos, Pepsi, Salsa2 Nachos, Coca-Cola, Salsa3 Nachos, Coca-Cola4 Nachos, Pepsi, Salsa5 Milk, Bread

Frequent Itemset Support

{Nachos} 80%{Salsa} 60%{Coca-Cola} 40%{Pepsi} 40%{Nachos, Salsa} 60%{Nachos, Coca-Cola} 40%{Nachos, Pepsi} 40%{Salsa, Pepsi} 40%


/ 15

Covering Set

CS(i ∣L) = {X ∈ L ∶ {i} ∪X ∈ L}

coverage(i ∣L) = ∣CS(i ∣L)∣

Frequent Itemset

{Nachos}{Salsa}{Coca-Cola}{Pepsi}{Nachos, Salsa}{Nachos, Coca-Cola}{Nachos, Pepsi}{Salsa, Pepsi}

i CS(i) coverage

{Nachos} {{Salsa}, {Coca-Cola}, {Pepsi}} 3{Salsa} {{Nachos}} 1{Coca-Cola} {{Nachos}} 1{Pepsi} {{Nachos}, {Salsa}} 2


/ 15

Substitutive Sets

A two-element itemset {x , y} is a substitutive itemset, if:

x ∈ L1,

y ∈ L1,

support({x} ∪ {y}) < ε, where ε is a user-defined thresholdrepresenting the highest amount of noise in the data allowed,∣CS(x ∣L)∩CS(y ∣L)∣

max{∣CS ∣L(x)∣,∣CS(y ∣L)∣} ⩾ mincommon.

i CS(i) coverage

{Nachos} {{Salsa}, {Coca-Cola}, {Pepsi}} 3{Salsa} {{Nachos}} 1{Coca-Cola} {{Nachos}} 1{Pepsi} {{Nachos}, {Salsa}} 2

∣CS(Pepsi)∩CS(Coca−Cola)∣max{∣CS(Pepsi)∣,∣CS(Coca−Cola)∣}

= 0.5


/ 15

Create Substitutive Sets RapidMiner operator


/ 15

Use Case: DBpedia

DBpedia knowledge base version 2014

sets of 3–item transactions {c1,p, c2}, where c1 and c2 classes ofsubject and object of RDF triple, and p property connecting s and o

SELECT ?c1 ?p ?c2WHERE {?s rdf:type dbpedia-owl:Organization .?s ?p ?o .?s rdf:type ?c1 .?o rdf:type ?c2 .FILTER(?p != dbpedia-owl:wikiPageWikiLink) .FILTER(?p != rdf:type) .FILTER(?p != dbpedia-owl:wikiPageExternalLink) .FILTER(?p != dbpedia-owl:wikiPageID) .FILTER(?p != dbpedia-owl:wikiPageInterLanguageLink) .FILTER(?p != dbpedia-owl:wikiPageLength) .FILTER(?p != dbpedia-owl:wikiPageOutDegree) .FILTER(?p != dbpedia-owl:wikiPageRedirects) .FILTER(?p != dbpedia-owl:wikiPageRevisionID)}


/ 15

Transaction generation

dbpedia'owl:MusicalAr2st44 dbpedia'owl:PopulatedPlace44

Norah_Jones44

Denton,_Texas4dbpedia'prop:origin4

Mark_Knopfler4 Gosforth4dbpedia'owl:hometown4

dbpedia'owl:MusicalAr2st44 dbpedia'owl:PopulatedPlace44

s4

s4 p4

p4 o4

o4

c14

c14

c24

c24

Transactions

{c1 dbpedia-owl:MusicalArtist, dbpedia-owl:hometown, c2 dbpedia-owl:PopulatedPlace }{c1 dbpedia-owl:MusicalArtist, dbpedia-prop:origin , c2 dbpedia-owl:PopulatedPlace }...


/ 15

Experimental Setup

FP-Growth: min number of itemsets = 500, max number of retries =15, min support: 1.0E-4),

Create Substitutive Sets: min support = 1.0E-4, min common= 0.7, epsilon =1.0E-5,

a sample of 100k results per each query,

desktop computer with 12GB RAM and CPU Intel(R) Core(TM)i5-4570 3.20GHz,

a single run of mining substitutive sets (for a single class and 100ktransactions) took several seconds on average (ranging from 2s to12s)


/ 15

Sample substitutive properties for the class Organisation

Item X Item Y Common

dbpprop:parentOrganization dbo:parentOrganisation 1.000dbpprop:owner dbo:owner 1.000dbpprop:origin dbo:hometown 1.000dbpprop:headquarters dbpprop:parentOrganization 1.000dbpprop:formerAffiliations dbo:formerBroadcastNetwork 1.000dbo:product dbpprop:products 1.000dbpprop:keyPeople dbo:keyPerson 0.910dbpprop:commandStructure dbpprop:branch 0.857dbo:schoolPatron dbo:foundedBy 0.835dbpprop:notableCommanders dbo:notableCommander 0.824dbo:recordLabel dbpprop:label 0.803dbo:headquarter dbo:locationCountry 0.803dbpprop:country dbo:state 0.753


/ 15

Summary

Introduced a model for substitutive itemsets mining

Preliminary tests of this model within the task of deduplication ofobject properties in an RDF dataset (DBpedia)


/ 15

Acknowledgements

Foundation for Polish Science under the POMOST programme,cofinanced from European Union, Regional Development Fund (NoPOMOST/2013-7/8) (2013-2015)

EU FP7 ICT-2007.4.4 (No 231519) ”e-LICO: An e-Laboratory forInterdisciplinary Collaborative Research in Data Mining andData-Intensive Science” (2009-2012)

Thanks to Ewa Kowalczuk for debugging the RapidMiner plugin


/ 15