markov logic network - university of illinoisswoh.web.engr.illinois.edu › ... › handout ›...
TRANSCRIPT
MarkovLogicNetworkMatthewRichardson
andPedroDomingos
IE5982016,PresentbyHaoWu(haowu4)
Motivation- UnifyingLogicandProbability
• Logicandprobabilityaretwomostimportwayofreasoning.• “Classic”AIfavorslogicapproaches,whichismostlyrulebased.
• Theoremproofing.• Cannotdealwithuncertainty,verylimitedsuccess.
• “Modern”AIapproachesaredominatedbymoreprobabilisticmethods,whichhandlestheuncertaintyandnoiseinrealdata.
• DeepLearning,PGMandetc.• Hugesuccess
• SowhywestillwanttohaveLogic?(Whynotlearneverything?)
Whylogicisstillinteresting
• Logic,especially,First-orderlogicprovideaexpressive,compactandelegantwaytoexpressknowledge.
• Itonlytake30+linetowritedowntheruleofSudokuinProlog(andthesamecodecanalsosolveit).Howmanydatadoyouneedtolearneverythingfromscratch?
• Wewantanicewaytorepresentandsolveourproblems(efficiently).• Useexpertknowledgetohelpthedatadrivensystem.
• MarkovLogicisawaytoconnectsLogicandProbability.• Logichandlescomplexity.• Probabilityhandlesuncertainty.
Background:MarkovNetwork
• Potentialfunctionsdefinedovercliques
∏Φ=c
cc xZxP )(1)(
Smoking
Cough
Cancer
Asthma
FirstOrderLogic• Constants,variables,functions,predicatesE.g.:Anna,x,MotherOf(x),Friends(x,y)
• Literal: Predicateoritsnegation• Clause: Disjunctionofliterals• Grounding: ReplaceallvariablesbyconstantsE.g.:Friends(Anna,Bob)
• World (model,interpretation):Assignmentoftruthvaluestoallgroundpredicates
Comparision
( ))()(),(,)()(
ySmokesxSmokesyxFriendsyxxCancerxSmokesx
⇔⇒∀
⇒∀FOL:
MRF:
Smokes(B)
Cancer(A) Cancer(B)
Friends(A,B)
Smokes(A)
MarkovLogicNetwork
• AMarkovLogicNetwork(MLN) isasetofpairs(F,w) where• F isaformulainfirst-orderlogic• w isarealnumber
*Andweneedadatabasethatcontainsconstantsforgrounding.
( ))()(),(,)()(
ySmokesxSmokesyxFriendsyxxCancerxSmokesx
⇔⇒∀
⇒∀
1.15.1
( ))()(),(,)()(
ySmokesxSmokesyxFriendsyxxCancerxSmokesx
⇔⇒∀
⇒∀
1.15.1
+ Two constants: Anna (A) and Bob (B)
Cancer(A)
Smokes(A)Friends(A,A)
Friends(B,A)
Smokes(B)
Friends(A,B)
Cancer(B)
Friends(B,B)
MarkovLogicNetwork:Definition• Eachgroundformuladefinesaclique
• isthenumberoftruegroundingofformulai• isthestate(truthvalue)ofatomsinformulai
MarkovLogicNetworks
• A template forgroundMarkovRandomField.• CanhavetypetoreducethenumberofpredicateXconstants.
• i.e.Humancanonlybefriendwithanotherhuman.
• Expressivity:• Whensetallweighttoinfinitelarge,itbecomesFOL.• Everyprobabilitydistributionoverdiscreteorfinite- precisionnumericvariablescanberepresented asaMarkovlogicnetwork.
Inference(SameasinferenceonMRF*)*SometimeneedalittletwistforMCMCstyleinference
• MAPInference:
• ConditionalInference
)|(maxarg xyPy
• Learnfromadatabase• Cantolearnbothweights(parameters)andFOLformula(structure):
• Learningweights.• Byoptimizelikelihood.
• Learningformula:(InductiveLogicProgramming)• AnILPsystemwillderiveahypothesised logicprogramwhichentailsallthepositiveandnoneofthenegativeexamples.
• UseexistingInductivelogicprogrammingsystem.
Learning
Learningweight
• Optimizelikelihood.(Generativeapproach)
• Generalizedtoohard,doPseudo-likelihoodinstead.• CountingtruegroundingsofafirstorderclauseinaKBis#Pcomplete
• Optimizeconditionallikelihood.(Discriminativeapproach)
( )∑ ∑
∑
=
−====
'
),'(exp
log),()|(log)(
yi iix
xi
ii
xynwZ
ZxynwxXyYPwf
( )∑ ∑
∑
=
−===
xi ii
iii
xnwZ
ZxnwxXPwf
)'(exp
log)()(log)(
∑ ==l
lll xMBxXPxPL ))(|(log)(log
Application- Entityresolution(CitationDB)
• Author(bib,author)Title(bib,title)Venue(bib,venue)• HasWord(author,word)• HasWord(title,word)• HasWord(venue,word)• SameAuthor(author1,author2)• SameTitle(title1,title2)• SameVenue(venue1,venue2)• SameBib(bib1,bib2)
Application- Entityresolution
• Title(b1,t1)∧ Title(b2,t2)∧ HasWord(t1,+w)∧ HasWord(t2,+w)⇒SameBib(b1,b2)
• Author(b1,a1)∧ Author(b2,a2)∧ SameBib(b1,b2)⇒SameAuthor(a1,a2)
• Author(b1,a1)∧ Author(b2,a2)∧ SameAuthor(a1,a2)⇒Samebib(b1,b2)