Logo
Data Snapshot
Login Register
wilson-facct21
auto-generated 27-Oct-2021
Building and Auditing Fair Algorithms: A Case Study in Candidate Screening Christo Wilson Northeastern University cbw@ccs.neu.edu Alan Mislove Northeastern University amislove@ccs.neu.edu Avijit Ghosh Northeastern University avijit@ccs.neu.edu Lewis Baker pymetrics, inc. lewis@pymetrics.com Shan Jiang Northeastern University sjiang@ccs.neu.edu Janelle Szary pymetrics, inc. janelle@pymetrics.com Kelly Trindel pymetrics, inc. kelly@pymetrics.com Frida Polli pymetrics, inc. frida.polli@pymetrics.co

Keywords

code for their candidate groups being the eeoc discrimination that may impact group had corresponding demographic candidates to their clients process and no model data from these games data from four clients model for given client

Sources: Brookings

Pathway Demo and Download Data

You can download the excel file containing Categories, Goals, Measures and Actions. Please use Lunge App, to try the pathway data.
Download Pathway Data
This section lists the Pathway type, Priority/ function areas, Sectors/ fields of life and People/ roles for the pathway.
Type:   Expert Advice & Recommendations
Priority/ Function Areas:   Business & Industrial  Finance  People & Society
Sectors/ Fields of Life:   Economy-wide
People/ Roles:   Organizations  Multi-Nationals  Governments
This section lists the Categories, Goals, Measures and Actions included in the pathway.
Categories:   values   candidate   audit   would   christo   given   black   canada   company   group   model   fairness   virtual   pymetrics   adverse   impact   cooperative   algorithm   testing   feature   study   seeker   machine   wilson   result
Profiles:   this games   this properties   an games   this questions   the model   we question   the results   any findings   canada manuals   the study   the groups   the data   the process   the agreement   a client   the values   the categories   a process   a heldout   all parties   the rate   datum scientists   we model   the analyst   each distribution   the description   the way   this pymetrics   another issue   polli building   the choice   the ratio   this dataset   this documents   a audit   the facets   this results   a pymetrics   the client   its system   the issues   the focus   the developer   the contract   the notebooks   its models   all paths   this demonstration   this model   that capabilities   that audits   the team   this categorization   the distributions   the class   the fourfifths   the source   the dataset   this experiences   the transparency   the search   the audit   blackbox systems   the imputer   we audit   the privilege   the threshold   this process   the group   the subject   a scientist   the models   the concerns   all models   that pymetrics   the doctrine   we attack   the job   this tests   the modules   this datasets   the party   that players   the deliverable   the auditors   this tools   pymetrics knowledge   its process   we testing   the systems   this voices   an groups   the goal   the datasets   a player   this audit   this stickers   proce pymetrics   a model   the company   california law   every company   the scientist   a employer   an audits   the metrics   its systems   the notebook   a employee   this actions   the code   inc measures   the assessment   we results   this assumption   fridapollipymetricscom activists   the axis   this choices   the public   a part   its choice
Goals:   assume following model   produce gameplay data   impute missing values   reconstruct incumbent teria   scale numerical data   customize adverse testing   test next step   cause applicant pool   quantify investigating bias   take optional survey   cover following aspects   test called impact   outline uses startup   copy numerical data   provide mapped dataset   question use choice   train using models   require specific permission   bless biased practices   release sound assumptions   use separate testing   7using andersondarling test   offer reasonable level   scoping employment selection   survey startups information   use proprietary implementation   investigate concepts race   use evaluation datasets   disclose demographic data   contain science biases   investigate used sources   present used process   investigate found questions   fail adverse tests   refer case study   document evaluated practices   face malicious company   cover potential cases   develop security gram   suit particular client   ing operationalize choice   conduct agreed audit   use bias ratio   examine disparate treatment   ing child services   improve overall application   auto adverse tests   use ratio metric   avoid adverse tests   define frame work   give unique challenges   audit candidate tool   implement stated guarantees   retain editorial discretion   draw direct line   sign agreeing contract   support unique needs   use imputed datasets   perform backtesting impact   begin new engagement   include ask questions   audit audience bias   ethnicityrace available categories   provide processing power   present technical overview   replace feature values   address related questions   fail human oversight   introduce cooperative audit   alter adverse assessment   perform attempt tests   abandon training process   monitor border personalization   discuss control flow   implement ing product   play pymetrics games   deploy performant model   benefit research community   clean involves data   involve outsider audits   compare collected data   examine cybersecurity posture   include safe guards   undergo security audits   set unrealistic precedent   involve raises intervention   produce went model   asses mental qualities   assume auto priori   reflect case study   run rics code   offer additional suite   offer candidate service   scale differential impact   give remediate days   choose existing objective   undermine bias group   evaluate included models   underlie process pymetrics   share proprietary information   test imputed datasets   cover code data   evaluate practices claims   cover wide range   retain speak right   elicit given behaviors   exhibit gameplay characteristics   undertake steps pymetrics   promote adopted properties   write significant deviations   use psychological surement   release pymetrics process   encompass template notebook   review key aspects   implement source consideration   involve use choices   complete includes checklist   follow contracts 46   present audited process   include demographic data   deploy biased model   raise case study   raise building algorithms   implement impact testing   debias meant malicious   perform impact testing   use nonparametric mannwhitney   set new precedent   include equal proportion   deploy bestperforming model   reevaluated original model   observe live demonstration   provide adverse framework   thank anonymous reviewers   audit adverse impact   understand target role   ing case study   complete given audit   port available information   learn science pipeline   study job performance   dictate conducted tests   duct agreed audit   implement fourfifths rule   use internal tools   implement scientistfacing process   influence adverse assessment   match audited one   customize model training   want motivations things   test following impact   complete pymetrics games   support significant amount   answer examined questions   examine source code   find alternate model   alter audit methods   isolate hires behaviors   use outcomebased model   posttraining impact testing   give performance data   undergo independent audit   outperform original model   know evaluated manner   select performant model   mitigate evaluating bias   use bias group   use case study   perform internal study   give identifying information   clean feature values   investigate legal space   motivate inves tigation   construct group datasets   give privileged access   evaluate training portion   approximate potential pool   release biased model   investigate single company   perform adverse testing   answer posed questions   use proprietary tool   use source code   use updated data   circumvent fairness checks   audit screening product   produce biased model
Measures:   fair classification   sound assumptions   impact assessment   different characteristics   technical assistance   human behavior   legal limitations   comparative study   50th percentiles   performance requirements   third metrics   small part   demographic labels   update data   similar attributes   essential principle   internal tools   pymetrics tool   unique challenges   cybersecurity posture   out group   job candidates   bestperforming model   bend employees   hispanic twoormore   gamified surement   malicious insiders   good datasets   sociotechnical pymetrics   pymetrics code   pymetrics suite   brief introduction   partisan bias   complex processes   outcomebased model   overall implementation   significant amount   several points   computational power   more discussion   independent contractors   challenge questions   new products   follow questions   white twoormore   internal example   fair algorithms   demographic groups   numerous ples   source code   safe guards   long standing   50th percentile   widespread problem   predictive analytics   bias practices   major groups   female twoormore   linger questions   mobile app   legal space   intentional malfeasance   psychological studies   arbitrary data   institutional oversight   confidential games   due differences   traditional boards   empirical analysis   regulatory compliance   financial means   accountability gap   private companies   process pymetrics   third person   general regulation   regulatory limitations   broad freedoms   real conflicts   transparent researchers   cooperative auditing   model predictions   confidential study   result pages   own expense   many choices   pymetrics standards   accredit body   malicious behaviors   algorithmic opacity   future audits   standard deviations   digital copies   predictive model   screen product   large population   thirdparty audits   specific details   demographic data   cooperative mode   give client   internal manuals   good incentive   give engagement   significant deviations   right questions   multiple times   anal ysis   virtual chine   mea surement   many academics   detail occupations   humanlike biases   follow model   bias group   specific criteria   predic model   initial onboarding   human capabilities   algorithm audit   critical concerns   specific audit   give group   feature values   many techniques   pymetrics games   andersondarling test   graphic survey   misin formation   reasonable mitigation   gamebased candidate   great potential   gross negligence   frame work   state guarantees   applicant data   enor number   correct issues   total template   minor groups   incumbent employees   ratio metric   fourfifths rule   homogeneous group   direct participants   correspond sessions   particular client   impact framework   performant model   transparent systems   open question   manual process   compa nies   unlabeled data   main concern   semiannual audits   geographic lines   process power   only inputs   adverse testing   train process   front pymetrics   employment selection   insider employees   penetration tests   devah pager   virtual machine   safety systems   deploy systems   available information   grow body   manual review   programmatic constraints   black twoormore   client engagements   additional suite   malicious client   aggregate information   legal dispute   datadriven discrimination   great extent   prior clients   editorial discretion   exist objective   search bias   pymetrics concerns   different groups   candidate service   first choice   potential seekers   top stories   evaluation datasets   adverse fairness   candidate tool   possible extent   technical correctness   significant instances   target role   intersectional fairness   specific permission   human error   male twoormore   sixth book   inter stage   internal study   sy tems   future acknowledgments   demographic information   new engagement   potential cases   source consideration   independent investigators   overt proxies   technical brief   event issue   minimum ratios   pymetrics choice   differential impact   overall application   malicious behavior   step pymetrics   scientistfacing process   compliant model   complete notebooks   heldout sets   original model   commercial gender   direct discrimination   annual meeting   online maps   complete engagement   complete engagements   language processing   significant impact   assume model   default codebase   social biases   broad occupations   predictive models   ethical demands   correspond jobs   equal opportunity   model performance   unique needs   conceptual space   exclusionary precedent   accurate algorithms   heldout testing   hire domain   miss data   additional product   hispanic asian   screen tool   social facets   gameplay characteristics   official blog   exist data   general questions   confidential brief   case study   demographic categories   quillian pager   male asian   minimum requirements   demographic pymetrics   second pymetrics   demographic survey   such audit   fabricate data   multiple building   artificial consensus   open source   demographic disclosure   security audits   intentional behavior   personal use   optional survey   different outcomes   miss game   baseline requirements   appbased platforms   label data   summary data   ing data   internal actors   separate testing   jupyter notebook   future pymetrics   additional filters   facial recognition   erroneous models   objective function   cific references   mous number   adverse assessment   assessment systems   selfreported data   pymetrics posture   pymetrics compliance   recommend tier   second series   algorithm audits   miss feature   reasonable choice   relevant regulators   numerical games   disparate treatment   online adver   train models   algorithmic logic   border personalization   protect groups   several tools   complex task   nonparametric mannwhitney   next question   predictive permutation   confidential methods   fair model   correspond information   blank template   productive inquiry   mandatory audits   choose threshold   future work   core pymetrics   black males   inherent tradeoffs   more companies   international personalization   screen service   extra bulwark   inves tigation   consistent manner   prior permission   associate modules   substantive problems   custom module   direct line   characteris tics   tive relationship   relevant agencies   proprietary code   threat model   template notebook   nonparametric tests   asian twoormore   confidential presentation   important background   active engagements   own set   specific motivations   black asian   impact testing   exist proposals   protect attributes   human intervention   extensive taxonomies   diverse set   longitudinal analysis   prior applicants   new precedent   international organization   serious projects   specific ways   virtual environment   posttraining testing   sociotechnical background   imaginary employees   client engagement   unprecedented level   confidential fairness   many performance   principled consensus   private company   second scientist   equal proportion   second guards   adverse tests   malicious company   neutral sources   debiase circumvention   fair determination   many objectives   represen tation   external team   bias ratios   adverse impact   correspond categories   direct payment   peerreviewed studies   step process   good faith   nical correctness   acceptable range   extraordinary access   disparate impact   plural person   give context   proprietary implementation   advertise interface   correspond notebook   recent engagements   human beings   external auditors   intersectional groups   hire practices   research community   statistical analysis   demographic subgroups   legal doctrine   sincere hope   careful protocols   investigative secrecy   fairness issues   recursive interrogation   outsider audits   consistent code   practical testing   preemployment assessment   incum employees   audit methods   human risks   applicant pool   online reviews   second question   specific clients   use cases   racial discrimination   human job   perform employees   first person   extensive changes   exist regulations   statistical testing   publiclydocumented audit   high numbers   standard training   final models   key aspects   standard process   modern systems   websci dino   racial groups   regulate area   science pipeline   kruskal tests7   salient data   sociotechnical systems   analysis consideration   political searches   virtual event   metric ratio   full transparency   miss games   business practices   willing company   protective services   cognitive characteristics   pro cess   associate data   tive model   tected groups   deploy models   broad classes   more audits   unify approach   search questions   compliant gram   price discrimination   cal space   social media   social accountability   specific operations   code data   first issue   anonymous reviewers   reas surance   develop software   demographic group   demographic games   incum bents   power imbalances   particular job   human oversight   erative audits   fair sets   online evidence   leverage source   online platforms   game traits   request permissions   accept bestpractices   legitimate concerns   audit interrogation   differential validity   fairness guarantees   candidate product   source data   candidate screening   gameplay data   proprietary information   ethical norms   tive clients   specific client   sociotechnical guards   digital journalism   pymet tool   clear aheadoftime   algorithm auditing   percentile thresholds   third party   human biases   model training   stringent thresholds   several deviations   concern guards   templated model   job seekers   vary algorithms   independent experts   impact tests   academic community   control flow   plete information   different context   automate models   available categories   major university   hierarchical taxonomy   independent audit   test framework   artifact collection   willing audits   internal pymetrics   previous section   classic design   wide range   verge https   unequal tation   potential highperformers   great care   narrow confines   recent engagement   substantive issues   predictive performance9   specific questions   performance data   few examples   natural language   first copyrights   fake data   numerical data   representative datasets   bias permutation   armslength relationship   canada issue   train model   potential pool   algorithmic auditing   datum scientist   single company   sound fairness   full citation   mlbased hiring   reasonable level   noncustomized model   algorithmic pricing   scientific study   geted role   confidential code   accuracy disparities   interpretable models   predictive performance   public largerscale   mean values   ing product   ethnographic study   significant numbers   algorithmic fairness   job performance   intersectional disparities   few audits   psychological surement   specific results   geographic distribution   technical overview   empty value   little variability   external audit   enough information   third pymetrics   fairnessenhancing interventions   artificial intelligence   equitable outcomes   final discretion   black boxes   impute dataset   increase adoption   give figure   international conference   cooperative audits   positive audit   complete notebook   group datasets   key precept   investigative journalists   different behaviors   minimum ratio   helpful guidance   public perspective   probono audit   70th percentile   digital dominance   insider attacks   potential behaviors   automate systems   politicallyrelated engine   internal audit   federal regulations   debiase techniques   highperforming bents   fairness claims   intrinsic qualities   unfair model   objec relationship   pro gram   online delivery   regulatory demands   mlbased systems   median tion   longitudinal pymetrics   asian females   treatment disparity   target advertising   standard practices   alternate model   heldout set   pymetrics process   customize products   clever users   follow aspects   unrealistic precedent   more imputation   new models   predictive requirements   scaffold code   multiple fairness   cooperative format   present pymetrics   algorithmic curation   foremost issues   live demonstration   replicable framework   actual systems   several people   adverse process   volving algorithms   lease permutation   cooperative audit   adverse metrics   potential objectives   use datasets   audience bias   military operations   supply pymetrics   demographic characteristics   bias ratio   acceptable bounds   absolute terms   miss values   give data   identify information   adverse framework   sufficient guards   substantive impact   fairness checks   median values   female asian   bias model   next step   past pymetrics   white asian   oversee scientist   median imputation   privilege access   algorithmic accountability   online sources   ugesp fourfifths   mental qualities   proprietary tool   individual players   online markets
Actions:   examine requirements   debias malicious   challenge pymetrics   identify code   use study   include questions   require highlevel   draw distinction   receive data   investigate questions   involve employees   offer level   apply ml   offer suite   contain information   raise algorithms   examine logic   evaluate seekers   evaluate portion   isolate behaviors   complete checklist   reflect views   mis problems   remove biases   audit partisanship   close gap   liken auditing   raise study   undermine group   define work   replace scale   compare data   sign contract   meet standard   ensure correctness   describe cess   meet guarantees   assume model   copy data   review aspects   provide reassurance   mis addition   retain right   develop process   supply possibility   uncover biases   use mannwhitney   cover coverage   retain discretion   investigate soundness   mitigate concerns   check it   include guards   inform choice   rule representativeness   auto tests   inform pymetrics   upload it   consider ir   require disparity   use ml   give data   implement process   use surement   adopt policy   monitor personalization   recommend candidates   highscoring seekers   ing intersectionality   fear loss   miss traits   use dataset   offer systems   debiase circumvention   preserve compliance   answer questions   insure independence   remove outliers   navigate questions   evaluate aspects   adopt process   manipulate the   concretize assumptions   package model   obviate concerns   abandon process   evaluate fairness   ridesharing reviews   give information   define scope   use code   give opportunity   train 5note   meet standards   posttraining testing   adopt stance   reverse engineering   use machine   give days   interpret predictions   request changes   skip game   receive that   differentiate s   audit the   set precedent   minimize harm   apply models   operate raghavan   face company   select model   motivate tigation   identify issue   pas fairness   circumvent checks   discuss flow   raise issue   contain biases   identify features   elicit behaviors   produce models   structure audit   test this   maintain relationship   reflect tics   provide copy   find model   apply filters   deploy facct   incorporate tems   calculate metrics   provide bulwark   denote which   disseminate limitations   solicit data   underlie pymetrics   test impact   make copies   ask questions   provide surance   make results   implement consideration   complete audit   demonstrate outcome   tising maps   privilege compliance   audit tool   audit algorithms   support amount   cerns composition   provide dataset   analyse data   contain users   raise alarms   learn pipeline   provide us   identify highperformers   examine treatment   deliver findings   conduct attacks   disclose data   undergo audits   audit systems   surface information   convert concerns   use measures   reflect study   pas checks   assume priori   must 1   encourage pymetrics   present context   train replacement   employ data   rely ratio   concretize values   miss values   use tool   decommissions model   want things   reimagining codebase   introduce audit   map engagements   play game   ensure compliance   influence assessment   produce manuscript   scale 531   learn 13   test datasets   involve audits   clean data   use metric   divulge information   replace values   use datasets   match one   support needs   focus clusively   give freedoms   challenge work   investigate race   engage experts   suit client   see 54   apply learning   detect discrimination   ing algorithms   involve cooperation   debiase cess   use what   suit needs   contextualize audit   prepare it   assess efficacy   scale data   maintain posture   subvert guarantees   establish number   avoid treatment   measure capabilities   conduct audits   gamebased screening   offer service   approach issues   refer study   find issues   examine guards   release model   audit impact   investigate sources   alter methods   raise complication   accept remuneration   produce number   perform study   jeopardize privacy   put access   reevaluated ratios   interrogate processes   onet occu   control size   scoping questions   cover aspects   fail tests   remediate them   complete survey   facilitate diting   review documents   mislead 42   develop model   develop protocols   evaluate characteristics   highperforming employees   identify issues   use that   design audit   accu racy   vote power   seek ers   clean values   obey requirements   audit interface   give copies   rule ajunwa   run code   thank reviewers   audit pages   perform tests   use models   identify instances   benefit community   fail oversight   explain predictions   input it   measure discrimination   mizing performance   run audit   take collusion   audit what   see 433   use tools   structure audits   differentiate bents   train beings   face eg   examine claims   serve significance   honore requests   show results   examine tion   outperform model   extract information   preserve secrecy   discuss questions   perform testing   provide framework   provide 10onet   scale impact   dictate tests   map practices   establish precedent   make criteria   cover 16   outline process   highperforming information   involve correct   remove references   encompass notebook   disseminate methods   outline startup   test step   trust you   ethnicityrace categories   asses qualities   maintain database   use group   release process   present process   select that   avoid conflicts   use set   force scientist   see collection   pas them   shed light   include malefemale   quantify bias   doubt sincerity   identify that   play games   7using test   prevent pymetrics   examine posture   contain data   approximate pool   include audits   perform analysis   bear notice   include proportion   examine classes   use implementation   organize discussion   support themselves   compare mance   conduct audit   set values   require permission   raise bar   train models   retain remuneration   preserve objectivity   construct datasets   evaluate that   filter outliers   emulate human   choose objective   express interest   screen objective   generate reports   take place   know manner   produce model   use testing   map these   audit bias   urge nies   question choice   involve intervention   present casestudy   provide labels   prevent impact   preexist imbalances   assess variety   approach question   impact guarantees   reveal demographics   raise concerns   develop gram   achieve goal   investigate company   scoped audit   pas rates   use model   outline framework   motivate series   observe demonstration   measure formation   examine code   reveal focus   proach issue   cover all   customize testing   alter assessment   introduce designs   study performance   cover concerns   raise issues   prevent scientist   precede audit   create systems   see adoption   investigate space   ing choice   ensure independence   produce outcomes   investigate pymetrics   evaluate models   construct sets   imized harm   use ie   motivate need   offer number   investigate ability   audit process   involve choices   document practices   include data   provide data   measure personalization   offer design   impute values   confine activities   introduce that   expand choice   see 3   cover people   cause pool   train goal   algorithm audits   produce data   improve application   give access   ask people   bless practices   mitigate s   register 166   supply that   undergo audit   pass audit   impact composition   select employees   abandon them   implement guarantees   assess fit   spend day   share information   justify work   contain pii   add layer   survey information   observe variability   compound biases   begin engagement   produce facct   start games   incentivize companies   provide power   skip games   evaluate suitability   implement gorithms   take care   understand role   impact players   avoid tests   ing services   investigate instances   expand audit   scoping selection   give challenges   draw line   show change   give notebooks   port information   implement testing   evaluate claims   use extent   perform impact   take survey   asses performance   promote properties   develop models   tweak models   maintain security   reconstruct teria   reevaluated model   ask scientist   prepare data   follow 46   grant compromise   deploy systems   undertake steps   require assistance   serve grants   use 50   perform audit   ask do   ing study   complete probono   document audits   remove impact   ensure results   cover range   copy accuracy   mitigate bias   discuss design   correspond sandvig   construct baseline   produce impact   present results   investigate impact   audit product   address questions   present overview   cover cases   use data   audit point   require imputation   customize training   manage relationship   duct audit   conduct search   serve lundberg   accept payment   investigate effects   exhibit characteristics   cover data   deploy model   revolve issue   reflect dataset   release assumptions   avoid impact   audit fairness   outline requirements   undertake pymetrics   motivate us   manipulate goal   include information   evaluate performance   contain results   reference ajunwa   surface applicants   show answer   play suite   implement rule   pretraining 19   write software   evaluate credibility   impact performance   replicate studies   complete games   map doctrine   discuss issues   make claims   provide guidance   commence audit   see testing   use ratio   perform annual   contain code   create grounds   ask players   write deviations   veloping framework   post jobs   present work   implement product   define process
This section lists the details about pathway-specific attributes.