Eswc2009

ReduCE Fanizzi, d'Amato, EspositoReduCE Fanizzi, d'Amato, Esposito

ReduCE: ReduCE: A Reduced Coulomb Energy Network Method A Reduced Coulomb Energy Network Method

for Approximate Classificationfor Approximate Classification

Dipartimento di InformaticaDipartimento di InformaticaUniversità degli studi di BariUniversità degli studi di Bari

Nicola Nicola FanizziFanizziClaudia Claudia d'Amatod'AmatoFloriana Floriana EspositoEsposito

ESWC 2009ESWC 2009 22ReduCE Fanizzi, d'Amato, EspositoReduCE Fanizzi, d'Amato, Esposito

Table of ContentsTable of Contents

MotivationMotivationLearning RCE Learning RCE NetworksNetworksApproximate Classifications of IndividualsApproximate Classifications of IndividualsExperimentsExperimentsConclusions & OutlookConclusions & Outlook


MotivationMotivation

Classic ML techniques for building inductive Classic ML techniques for building inductive classifiers for SW representations classifiers for SW representations explicitexplicit models: new concepts models: new concepts implicitimplicit models: neural networks, support vector models: neural networks, support vector

machines, graphic probabilistic modelsmachines, graphic probabilistic modelsInductive methods for classificationInductive methods for classification

often more often more efficientefficient and and noise-tolerantnoise-tolerant than standard than standard methodsmethods

enables enables approximationapproximationbetter exploitation of the inherently incomplete better exploitation of the inherently incomplete

information in Kbs for specific tasksinformation in Kbs for specific tasks


Applications of Inductive ModelsApplications of Inductive Models

Approximate instance-checkingApproximate instance-checkingThis can be also exploited forThis can be also exploited for

approximate retrievalapproximate retrievalsubsumptionsubsumption......

It also provides alternative methods for It also provides alternative methods for ontology populationontology population

Ultimately, may be used for completing ontologies Ultimately, may be used for completing ontologies with with probabilisticprobabilistic assertions assertions enabling more sophisticate approaches to dealing with enabling more sophisticate approaches to dealing with

uncertaintyuncertainty


Learning ProblemLearning Problem

Given a Given a target concepttarget conceptTrainTrain a model (hypothesis) a model (hypothesis) hhQQ using:using:

Set of pre-classified individuals: Set of pre-classified individuals: examplesexamplesA A knowledge baseknowledge base KK as background knowledge as background knowledge

thenthen

Use the learned model to classify all other Use the learned model to classify all other individuals:individuals:Given Given hhQQ and and xx00

Output Output hhQQ((xx00)) and possibly the likelihood of this assertionand possibly the likelihood of this assertion


Training SetTraining Set

A A limitedlimited number of individuals number of individuals for which the intended classification is knownfor which the intended classification is known

Hypothesis: Hypothesis: the function to be approximatedthe function to be approximated


The Inductive Model: RCE NetworksThe Inductive Model: RCE Networks

Q ¬Q

λ1 λ2 λ3 λN

x1 x2 xd

categorylayer

acj

pattern(radii)layer wjk

inputlayer

. . .

. . .


Training the ModelTraining the Model


RCE Model RCE Model ConstructionConstruction

▪ ▪▪

▪▪

▪

▪▪

▪

▪

▪▪

▪

▪

▪ ▪▪

▪

▪

▪

▪

11 22 33

44 55 66


Similarity MeasureSimilarity Measure

Generalization of a pseudo-distance Generalization of a pseudo-distance dd [ESWC2008][ESWC2008]


RCE RCE FinalFinal Model Model

Prototypes (examples)Prototypes (examples)


((VanillaVanilla) ) ClassificationClassification Procedure Procedure


ExtensionsExtensions

generalizing the generalizing the decision-makingdecision-making step: step:

likelihood:likelihood:


ExperimentsExperiments

For each ontologyFor each ontology100100 satisfiable query concepts randomly satisfiable query concepts randomly generated by composition (conjunction / disjunction) of generated by composition (conjunction / disjunction) of

(2 through 8) primitive and defined concepts(2 through 8) primitive and defined conceptsEvaluationEvaluation

comparing inductive responses comparing inductive responses to those returned by a standard reasoner (Pellet 2)to those returned by a standard reasoner (Pellet 2)

IndicesIndicesmatch ratematch rate: : identical classificationidentical classificationomission error rateomission error rate: : 0 vs. 0 vs. ±±11commission error ratecommission error rate: : +1 vs. -1 or -1 vs. +1+1 vs. -1 or -1 vs. +1induction rateinduction rate: : ±±1 vs. 01 vs. 0

Several runs: rates averaged according to the Several runs: rates averaged according to the 632+ bootstrap632+ bootstrap procedure procedure


Experiments: OntologiesExperiments: Ontologies

Ontologies employed in the experiments:Ontologies employed in the experiments:


OutcomesOutcomes


Outcomes / 2Outcomes / 2


Outcomes / 3Outcomes / 3


Conclusions & OutlookConclusions & Outlook

ML method transposed ML method transposed to SeWeb to SeWeb representationsrepresentations

Experiments: Experiments: good performancegood performanceHigh match rateHigh match rateLow induction rateLow induction rateSome omissionsSome omissionsVery limited Very limited

commissionscommissionsLow variance wrt to past Low variance wrt to past

inductive methodsinductive methods

ExtensionsExtensionsforce binary responseforce binary response

Expected to augment Expected to augment induction rateinduction rate

Pre-computation of Pre-computation of prototypical individuals:prototypical individuals:

Medoids from ClusteringMedoids from Clustering

Use likelihood for Use likelihood for rankingrankingadding probabilities to adding probabilities to

assertionsassertions


Questions ?Questions ?

For offline contacts:For offline contacts:Nicola Nicola FanizziFanizzi [email protected]@di.uniba.itClaudia Claudia d'Amatod'Amato [email protected]@di.uniba.itFloriana Floriana EspositoEsposito [email protected]@di.uniba.it

Other methods / systemsOther methods / systemshttp://lacam.di.uniba.it:8000/~nico/research/ontologymining.html

The EndThe End

Eswc2009

Technology

model rce model construction

indicesmatch rate

inductive classifiers

inductive responses

learned model

standard methods

model hypothesis hq

commission error rate