UnBias: Emancipating Users Against Algorithmic...
Transcript of UnBias: Emancipating Users Against Algorithmic...
UnBias: Emancipating Users Against Algorithmic Biases for a
Trusted Digital Economy
Michael Rovatsos reusing lots of slides byAnsgar Koene, University of Nottingham
InformationservicesFreetouse→ nocompetitiononprice
Plentyofassets→nocompetitiononquantity
Competitiononqualityofservice
Qualitydeterminedbyrelevance
Filteringbecomeskey
Conveniencevs.choice
SocialnetworksUserexperiencesatisfactiononsocialnetworksites
Filter
Filtering
Propagationofinfluence
RecommendationContent-based– similaritems tothoseuserlikedCollaborative– whatsimilarusers likedCommunity– whatpeopleinsamesocialnetwork liked
Theroleofalgorithms
Userunderstandingofsocialmediaalgorithms
More than 60% of Facebook users are entirely unaware of any algorithmic curation on Facebook at all: “They believed every single story from their friends and followed pages appeared in their news feed”.
CHI 2015
RevealingNewsFeedbehaviour
Awarenessthroughexperimentation
Garbagein,garbageout?perpetuatingthestatus-quo
Manipulation: conflict of interest
TherisksofalgorithmicbiasInformationfiltering,orranking,implicitlybiaseschoicebehaviour
Manyonlineinformationservicesarepaidforbyadvertisingrevenue
Conflictofinterest:promoteadvertisementvs.matchuserinterests
Advertising,whichinherentlyinvolvesmanipulation,becomesubiquitous
Personalizedfilteringcanalsobeuseforpoliticalspin/propagandaetc
TrendingTopicscontroversy
Q&AwithN.Lundblad(Google)“Humanattentionisthelimitedresourcethatservicesneedtocompetefor.Aslongasthereexistcompetingplatforms,lossofagencyduetoalgorithmsdecidingwhattoshowtousersisnotanissue.Userscanswitchtootherplatforms.”
Nicklas Lundblad,HeadofEMEAPublicPolicyandGovernmentRelationsatGoogle
Seealso:
2011FTCinvestigationofGoogleforsearchbias
EUcompetitionregulationvsGoogle
‘Equalopportunitybydesign’“Toavoidexacerbatingbiasesbyencodingthemintotechnologicalsystemsaprincipleof‘equalopportunitybydesign’—designingdatasystemsthatpromotefairnessandsafeguardagainstdiscriminationfromthefirststepoftheengineeringprocessandcontinuingthroughouttheirlifespan.”
BigData:AReportonAlgorithmicSystems,Opportunities,andCivilRights“,WhiteHousereportfocusedontheproblemofavoiding
discriminatoryoutcomes
TheUnBias ProjectWP1:Co-productionofcitizeneducationmaterialswithyoungpeople
WP2:Fairalgorithmsandtoolsforexposingalgorithmicbias
WP3:Qualitativeresearchintoofusers’sense-makingbehaviour
WP4:Developinganinformationandeducationgovernanceframework
Two-yearprojectfundedbyEPSRC(£1.1mbudget),collaborationbetweenNottingham(coordinator),Edinburgh,andOxford
Challengesrelatingtodatausedasinputstoanalgorithm
•Poorlyselecteddata,selectionbias
•Incomplete,incorrect,oroutdateddata
•Unintentionalperpetuationandpromotionofhistoricalbiases
Challengesrelatedtotheinnerworkingsofthealgorithmitself
•Poorlydesignedmatchingsystems
•Personalisationandrecommendationservicesthatnarrowinsteadofexpanduseroptions
•Decision-makingsystemsthatassumecorrelationnecessarilyimpliescausation
•Datasetsthatlackinformationordisproportionatelyrepresentcertainpopulations
Concernsregardingpersonalisation
•Socialconsequences:self-reinforcinginformationfiltering– the‘filterbubble’effect
•Privacy:personalisation involvesprofilingofindividualbehaviour/interests
•Agency:thefilteringalgorithmdecideswhichsegmentofavailableinformationtheusergetstosee
•Manipulation:people’sactions/choicesaredependontheinformationtheyareexposedto
Agency: user vs. algorithmChallengesrelatedtotransparency&accountabilityFilteralgorithmsprovidecompetitiveadvantage,detailsaboutthemareoftentrade-secrets
• Usersdon’tknowhowtheinformationtheyarepresentedwasselectedà norealinformedconsent
• Serviceusershaveno‘manual’overrideforthesettingsoftheinformationfilteringalgorithms
• Itisdifficultforserviceuserstoknowwhichinformationtheydon’tknowaboutbecauseitwasfiltered
Casestudy
FairnessintaskcompositionCombinatorialproblemofallocatinggroupsofuserstosharedtasks,wheretaskrequestscomefromusers◦ Hardconstraintsrestrictthegroupingsandtaskpropertiesthatcanberealised inprinciple
◦ Softconstraintsdeterminewhichcoalitionstructuresandtaskfeaturesarepreferredbysystemand/orusers
Examples:◦ Ridesharing wheredriverssharetheircarswithotherpassengersforaspecificjourney(ourvanillaexample)
◦ Meetingscheduling,citizensciencetasks,workforceshiftallocation,clinicalworkflowmanagement,etc etc
Composition
Users Tasks
Composition
Users Tasks
Composition
Users Tasks
DiversityintaskcompositionIntraditionalmechanismdesign,globalallocationsarecomputedgivenindividualpreferencesandglobalcriteria◦ E.g.socialwelfaremaximisation,Paretooptimality,strategy-proofness,envy-freenessetc.
Mechanismsareproposedthatprovablysatisfytheseproperties,solutioncanthereforebeimposedonusers
Diversityimpliesthatuserscannotreportallpreferences◦ Systemnevercapturesallrelevantdecisionvariables
◦ Solutionscannotbecomputed/consideredexhaustively
◦ Utilityofsolutionscannotbedeterminedbyusersapriori
FromtaskallocationtotaskrecommendationKeyproblems:
1.Howtocompute“optimal”sets ofsolutions
2.Howto influence users’choices
3. Howtolearn users’preferences
UserandsystemutilityUser'sutilityfunction dependsonuser’srequirementsandpreferences
Global(system)utilityfunctiondependsonsocialwelfareandmaximisingtaskcompletion
Learning
Constraints
MIPfirst
MIPothers
MIP*
Objective
Objective
Objective
Constraints
Constraints
Hard feasibility constraints
MIP*
MIPfirst
Computingallocations
InfluencingusersWewanttomodifyusers'utilityartificiallysothattheirchoicesleadtoafeasibleglobalsolution
• ExplicitApproaches:• Intervention• (Possible)futurereward
• ImplicitApproaches:• discounts• taxation
MIPfirst
MIPothers
MIP*
MIPfirst
Objective
Constraints
Noiseless and Constant Noise Models
Logit Model (also goes into objective function)
Sponsored Solution
Taxation
ConclusionsTaskrecommendationscenarioexposeschallengesrelatedtofairness
Whatdoesitmeanforalgorithmstobe“fair”?
Howcanwemaphumannotionstocomputationalmodels?
Computationallimitationsvs.societaldemands
Wheredotheresponsibilitieslie:algorithm,platform,orusers?
ThisismuchbiggerintermsofethicsofAIthan“killerrobots”or“singularity”!