Predicting the Winner of the Vice Presidential Debate with AI
Tonight is the televised vice presidential debate in the US. Tim Kaine and Mike Pence will for 90 minutes put their case before the American people and the world. My team at Fuzzy.ai thought it would be a good use of our platform to give an accurate prediction of the outcome (“winner”) of this debate.
It can be hard for AI to make accurate predictions in this kind of situation. A statistical model would typically use data from multiple vice presidential debates to make predictions about the current ones. Because Pence and Kaine have never had a vice presidential debate before, and probably never will again, a statistical model would compare against other candidates in other years.
The problem with this kind of model is that it requires finding the least common denominator in the different debates — which party won? How tall was the candidate who won? How old? From which state? — to make predictions about the current year. It assumes that this year is a lot like other election years — which I think most of us disagree with.
Fuzzy.ai is a tool for doing combined heuristic and statistical models. It can take human knowledge about a problem, encode it into a form that can be used in calculations, and then improve that knowledge over time based on feedback. We don’t have the ability to do a feedback process (again, just one debate) but we can build a heuristic model.
To develop the rules for the heuristic model, we’re using the pre-debate media coverage from a number of different news organizations: the New York Times, Washington Post, CNN, Politico, and Vox. We looked through the coverage to see what challenges the media thinks each VP candidate needs to do to win the debate.
We’ve taken these challenges and marked them up in our model as inputs. We’re considering these as the “IF” parts of a lot of “IF-THEN” statements. If Mike Pence can do this, then he will win the debate. If Tim Kaine can avoid doing that, then he will win the debate.
Here are the challenges we’ve taken out of these articles for Mike Pence:
Defend Donald Trump’s character and temperament
Defend Donald Trump’s tax returns (leaked this week)
Defend Donald Trump’s comments on Alicia Machado
Defend Donald Trump’s comments about Hillary Clinton’s marriage
Define his role as vice president
Put on a lively show
Avoid obvious mistakes
Make Trump seem normal
Attack the Clinton Foundation
Attack Clinton over Benghazi
Attack Clinton over her marriage to Bill Clinton
Share personal perspectives about Trump
Defend own position on abortion
Defend own position on gay rights
And here are the ones we see for Tim Kaine:
Define his role as vice president
Challenge Mike Pence on social conservative issues
Put on a lively show
Avoid obvious mistakes
Go on the attack
Associate Pence with Donald Trump’s excessive policies
Reassure progressives
Connect with young voters
Share personal perspectives about Clinton
Defend history of gifts as governor
Defend history as defense attorney for unsavory criminals
It should be clear that we’ll be looking at much more nuanced and detailed information on each candidate’s performance than just their height and party affiliation.
(Also, it should be clear that we don’t take a stance on whether either candidate should do any of these things. Some seem pretty dirty and underhanded. But these are the challenges that the media has set out as the necessary conditions for each candidate to “win”.)
After the debate, we’ll score the candidates’ performance on each of these challenges on a scale from 1 to 5, and we’ll use Fuzzy.ai to generate an overall score for the candidate in the debate.
We’ll then compare the results against polls that should come out in the next 24–48 hours about who “won” the debate, to see if this model was actually predictive.
