querying knowledge graphs by example en;ty...
TRANSCRIPT
UsabilityChallenges
QueryingKnowledgeGraphsbyExampleEn;tyTuplesNandishJayaram1+,ArijitKhan2*,ChengkaiLi1,XifengYan2,RamezElmasri1
UniversityofTexasatArlington1 UniversityofCalifornia,SantaBarbara2
RelatedWork
N.Jayaram,A.Khan,C.Li,X.YanandR.Elmasri.QueryingknowledgegraphsbyexampleenOtytuples,TKDE2015.N.Jayaram,M.Gupta,A.Khan,C.Li,X.YanandR.Elmasri.GQBE: Querying knowledge graphs by example enOty tuples,ICDE2014.DemoURL:hXp://idir.uta.edu/gqbe
TechnicalDetailsandDemo
QueryProcessing
1
10
100
1000
F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20Que
ryProcessingTime(secs.)
Query
GQBE NESS Baseline
12 13 18 10 8 10 8 12 8 8 11 9 7 11 8 9 9710 7#edgesinMQG
q Bigandcomplexdata:Lackofschema,challengingtousersanddevelopers.
q Howtoquerythegraph,andunderstandtheresults.
q Query-by-exampleinrelaOonaldatabases[Zloof’75].q Keyword search and keyword-based queryformulaOon[Changetal.’11].
q Set expansion [Wang et al.’07]. XML queryrelaxaOon[Amer-Yahiaetal.’2005].
q Exemplarqueries[Moinetal.,2014].
QueryInterface
Experiments
QueryGraphDiscovery
AnswerSpaceModeling
Mul;-TupleQueryGraphs
q Nodes(F)and(HL)aretwominimalquerytrees.q Node (F) corresponds to the sub-graph thatconnects Jerry Yang and Yahoo through edgefounded.
q Node (FGHLP) is theMQG, and it correspondstotheenOrequerygraphonthelem.
GroundtruthbasedaccuracycomparisonofGQBEandNESS.ComparisonofGQBE,NESSandExemplarQueries.Themeasuredparametersareprecision-at-k,MeanAveragePrecisionandnormalizedDiscountedCumula;veGain.
Pearson Correla;on Coefficient (PCC) between GQBE and AmazonMechanicalTurkWorkers, fork=30.MTurkworkerswerepresentedwithanswerpairsandaskedfortheirpreferencebetweenthetwoanswers ineachpair.20000suchopinionscollected.
Queryprocessing;mesofGQBE,NESSandBaseline.
Obtain neighborhoodgraph.
Weight edges usingheurisOcs and use agreedy approach too b t a i n a s m a l l e rconnectedMQG.
A l l the super -graphs of a nullnodearepruned.
(GHL), a node that does not haveanyanswers,isanullnode.
Ini;alLaXce
Lower Bound (LB): edgematching score of thecorrespondinggraph.
Upper Bound (UB): scoreof the highest-scoredsuper-graphinthelaice.
GQBEArchitecture
MulOple MQGs are mergedbased on edge labels andvertexlabelmatches.
Modified laXce withrecomputed upperboundary
KnowledgeGraphs
Input:anexampleofwhattheuserwantstofind.
AccuracyofGQBEonmul;-tuplequeries,k=25.
LaXcenodesevaluatedcomparison.
Queryprocessing;mefor2-tuplequeries.
ParOallyfundedby:+NowwithPivotal,*NowwithNanyangTechnologicalUniversity,Singapore
1.0E-2
1.0E-1
1.0E+0
1.0E+1
1.0E+2
1.0E+3
1.0E+4
Que
ryProcessingTime(secs.)
Combined(1,2)Tuple1+Tuple2
0
40
80
120
160
200
F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 F12 F13 F14 F15 F16 F17 F18 F19 F20
#ofLaX
ceNod
esEvaluated
Query
GQBE Baseline840
12 13 18 10 8 10 8 12 8 8 11 9 7 11 8 9 7 9710#edgesinMQG
DATASETS:Freebase(28Mnodes,47Medges,5400relaOonships)DBpedia(759Knodes,2.6Medges,9100relaOonships)QUERIES:20QueriesonFreebase8QueriesonDBpedia
Q u e r yGraph1
Q u e r yGraph2