S TATISTICAL F OUNDATIONS OF V IRTUAL D EMOCRACY Anson Kahng , Min - - PowerPoint PPT Presentation

▶

Nov 19, 2022 358 likes •478 views

S TATISTICAL F OUNDATIONS OF V IRTUAL D EMOCRACY Anson Kahng , Min Kyung Lee, Ritesh Noothigattu, Ariel Procaccia, and Alex Psomas ICML 2019 A UTOMATING E THICAL D ECISIONS Donors Recipients A UTOMATING E THICAL D ECISIONS Donors Recipients

SLIDE 1

STATISTICAL FOUNDATIONS

OF VIRTUAL DEMOCRACY

Anson Kahng, Min Kyung Lee, Ritesh Noothigattu, Ariel Procaccia, and Alex Psomas

ICML 2019

SLIDE 2

Donors Recipients

AUTOMATING ETHICAL DECISIONS

SLIDE 3

AUTOMATING ETHICAL DECISIONS

Donors Recipients How do you make this decision?  Which recipient deserves the food?

SLIDE 4

A MODEST PROPOSAL

Donor: Type of donation:

Ask participants to cast a vote every time a decision needs to be made

<latexit sha1_base64="+NPkOo4eOGyvmqjiQWhMTQZXq0=">AB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0m0oMeCF48VTFtoQ9lsN+3SzSbsToQS+hu8eFDEqz/Im/GbZuDtj4YeLw3w8y8MJXCoOt+O6WNza3tnfJuZW/4PCoenzSNkmGfdZIhPdDanhUijuo0DJu6nmNA4l74STu7nfeLaiEQ94jTlQUxHSkSCUbS3zcZY4Nqza27C5B14hWkBgVag+pXf5iwLOYKmaTG9Dw3xSCnGgWTfFbpZ4anlE3oiPcsVTmJsgXx87IhVWGJEq0LYVkof6eyGlszDQObWdMcWxWvbn4n9fLMLoNcqHSDLliy0VRJgkmZP45GQrNGcqpJZRpYW8lbEw1ZWjzqdgQvNWX10n7qu5d192HRq3ZLOIowxmcwyV4cANuIcW+MBAwDO8wpujnBfn3flYtpacYuYU/sD5/AHfcI62</latexit>
<latexit sha1_base64="+NPkOo4eOGyvmqjiQWhMTQZXq0=">AB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0m0oMeCF48VTFtoQ9lsN+3SzSbsToQS+hu8eFDEqz/Im/GbZuDtj4YeLw3w8y8MJXCoOt+O6WNza3tnfJuZW/4PCoenzSNkmGfdZIhPdDanhUijuo0DJu6nmNA4l74STu7nfeLaiEQ94jTlQUxHSkSCUbS3zcZY4Nqza27C5B14hWkBgVag+pXf5iwLOYKmaTG9Dw3xSCnGgWTfFbpZ4anlE3oiPcsVTmJsgXx87IhVWGJEq0LYVkof6eyGlszDQObWdMcWxWvbn4n9fLMLoNcqHSDLliy0VRJgkmZP45GQrNGcqpJZRpYW8lbEw1ZWjzqdgQvNWX10n7qu5d192HRq3ZLOIowxmcwyV4cANuIcW+MBAwDO8wpujnBfn3flYtpacYuYU/sD5/AHfcI62</latexit>
<latexit sha1_base64="+NPkOo4eOGyvmqjiQWhMTQZXq0=">AB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0m0oMeCF48VTFtoQ9lsN+3SzSbsToQS+hu8eFDEqz/Im/GbZuDtj4YeLw3w8y8MJXCoOt+O6WNza3tnfJuZW/4PCoenzSNkmGfdZIhPdDanhUijuo0DJu6nmNA4l74STu7nfeLaiEQ94jTlQUxHSkSCUbS3zcZY4Nqza27C5B14hWkBgVag+pXf5iwLOYKmaTG9Dw3xSCnGgWTfFbpZ4anlE3oiPcsVTmJsgXx87IhVWGJEq0LYVkof6eyGlszDQObWdMcWxWvbn4n9fLMLoNcqHSDLliy0VRJgkmZP45GQrNGcqpJZRpYW8lbEw1ZWjzqdgQvNWX10n7qu5d192HRq3ZLOIowxmcwyV4cANuIcW+MBAwDO8wpujnBfn3flYtpacYuYU/sD5/AHfcI62</latexit>
<latexit sha1_base64="+NPkOo4eOGyvmqjiQWhMTQZXq0=">AB7HicbVBNS8NAEJ3Ur1q/qh69LBbBU0m0oMeCF48VTFtoQ9lsN+3SzSbsToQS+hu8eFDEqz/Im/GbZuDtj4YeLw3w8y8MJXCoOt+O6WNza3tnfJuZW/4PCoenzSNkmGfdZIhPdDanhUijuo0DJu6nmNA4l74STu7nfeLaiEQ94jTlQUxHSkSCUbS3zcZY4Nqza27C5B14hWkBgVag+pXf5iwLOYKmaTG9Dw3xSCnGgWTfFbpZ4anlE3oiPcsVTmJsgXx87IhVWGJEq0LYVkof6eyGlszDQObWdMcWxWvbn4n9fLMLoNcqHSDLliy0VRJgkmZP45GQrNGcqpJZRpYW8lbEw1ZWjzqdgQvNWX10n7qu5d192HRq3ZLOIowxmcwyV4cANuIcW+MBAwDO8wpujnBfn3flYtpacYuYU/sD5/AHfcI62</latexit>

Issue: we must consult participants every time a donation occurs! Idea: what if we could predict how people would vote?

SLIDE 5

VIRTUAL DEMOCRACY

Data Collection Learning Aggregation “Learn models of people, and let the models vote”

SLIDE 6

DATA COLLECTION

Data Collection Learning Aggregation

Use features identified by Lee et al. (2017) to collect pairwise comparisons of potential recipients

SLIDE 7

LEARNING

Data Collection Learning Aggregation

Learn models of participants that capture their reported preferences on pairwise comparisons; let models vote

SLIDE 8

AGGREGATION

Data Collection Learning Aggregation

How do we aggregate these votes?

SLIDE 9

AGGREGATION

Fundamental question in virtual democracy: 

Which voting rule should we use to aggregate votes?

Desideratum: robustness to machine learning errors

We want voting rules that are likely to output the same result on both true underlying preferences and noisy votes

SLIDE 10

Theorem: Borda Count is robust under Mallows noise Theorem: PMC rules are not robust under Mallows noise

There always exists a profile with an acyclic pairwise majority graph, but whose noisy profile has an acyclic pairwise majority graph with a different topological ordering If the difference between the true Borda scores of two alternatives is small, then the probability that Borda swaps them in the noisy ranking is exponentially small

THEORETICAL RESULTS

SLIDE 11

Theorem: Borda Count is robust under Mallows noise Theorem: PMC rules are not robust under Mallows noise

“Don’t use PMC rules for virtual democracy” “Use Borda Count for virtual democracy”