基于图的人工智能基准测试框架 A Framework for Benchmarking Graph-Based Artificial Intelligence.pdf

ABSTRACT

TitleofThesis:AFRAMEWORKFORBENCHMARKING

GRAPH-BASEDARTIFICIALINTELLIGENCE

KentDanielO’Sullivan

MasterofScience,2024

ThesisDirectedby:ProfessorWilliamRegli

DepartmentofComputerScience

Graph-basedArtificialIntelligence(GraphAI)encompassesAIproblemsformulatedusing

graphs,operatingongraphs,orrelyingongraphstructuresforlearning.ContemporaryArtificial

Intelligence(AI)researchexploreshowstructuredknowledgefromgraphscanenhanceexisting

approachestomeettherealworld’sdemandsfortransparency,explainability,andperformance.

CharacterizingGraphAIperformanceischallengingbecausedifferentcombinationsofgraph

abstractions,representations,algorithms,andhardwareaccelerationtechniquescantriggerun-

predictablechangesinefficiency.AlthoughbenchmarksenabletestingdifferentGraphAIim-

plementations,mostcannotcurrentlycapturethecomplexinteractionbetweeneffectivenessand

efficiency,especiallyacrossdynamicandstructuredknowledgegraphs.

Thisworkproposesanempirical‘grey-box’approachtoGraphAIbenchmarking,providing

amethodthatenablesexperimentallytradingbetweeneffectivenessandefficiencyacrossdiffer-

entcombinationsofgraphabstractions,representations,algorithms,andhardwareaccelerators.

AsystematicliteraturereviewyieldsataxonomyofGraphAItasksandacollectionofintelligence

andsecurityproblemsthatinteractwithGraphAI.Thetaxonomyandproblemsurveyguidethe

developmentofaframeworkthatfusesempiricalcomputersciencewithconstrainttheoryinan

approachtobenchmarkingthatdoesnotrequireinvasiveworkloadanalysesorcodeinstrumen-

tation.

Weformalizeamethodologyfordevelopingproblem-centricGraphAIbenchmarksand

developatool