Email
Password
Remember meForgot password?
    Log in with Twitter

article imageUsing an AI named Polly to predict human behaviour

By Lisa Cumming     May 28, 2018 in Technology
AI is a game changer during election time and it's made its way into the Ontario election with Polly — the AI from Advanced Symbolics Inc., a company that describes itself at the intersection of artificial intelligence and market research.
Who is Polly?
Conceived in 2012, Polly is ASI's artificial intelligence that predicts human behaviour in many different instances.
"Polly is artificial intelligence and what she does is, in this case, goes out to all the ridings in Ontario and creates representative samples of each of those ridings so that she can accurately gauge the sentiment of that population and figure out who is likely to win the election," said CEO of Advanced Symbolics, Erin Kelly, about their AI — named Polly — that's currently being used to predict who is most likely to win the Ontario election.
How does Polly work?
Polly has also been working and analyzing over the course of the year, determining how people's minds — and votes — change when certain things happen. This technology, as Kelly said in this interview with Steve Paikin of The Agenda, is pulling from a much larger sample size (ridings and the popular vote) than a typical poll.
ICYMI: NDP momentum continues in Ontario election campaign
ASI has published Polly's methodology, which details how the AI operates. This sample size is pulled using Conditional Independence Coupling (CIC), which is an algorithm that goes around on online networks, cherry-picking "a sample of people that is representative of the overall population." According to ASI, the CIC can be restricted to sample from specific municipalities to broader countries and demographics within the sample populations collected are determined by other artificial intelligences.
"We've got over 100,000 people in this Ontario election poll now, and we are looking at every single one of them, everyday, seeing how they change," said Kelly.
CIC differs from traditional polling because no actual questions are asked to the people in the sample because, as ASI writes, "Asking questions may introduce bias by either forcing a person to take a stance on an issue they have previously been uninterested in, or biasing the response through wording of the question."
How do you create a sample without questions?
ASI uses two different artificial intelligences to replace the asking of questions: 'topic detection' and 'stance detection.'
Topic detection is the AI that determines if the person is discussing the topic in question— either negatively, neutrally or positively. To do this, the AI is given a set of articles that provide information pertaining to the topic in question and learns about the topic using the Latent Dirichlet Allocation (LDA) algorithm. This means the AI is learning about all the "arguments for and against, hot button items, and current events around the issue."
Stance detection is the AI that determines if the person, who's been discussing the topic in question, supports or opposes the topic.
"For each person in the sample, the AI examines all of their online communication within the study period to determine if it matches one of the learned topics. Based on this analysis, the AI assigns a probability if a person is discussing the topic or not. Stance detection is a separate AI that determines if a person opposes or supports the given topic." — From the briefer.
Am I in the sample?
ASI isn't doing anything that other companies haven't already done, they're mining from publicly-available data that you've already agreed to share.
"All mined data is taken from public online sites where each user has agreed to allow open access to their accounts and data. Mining of the data is done in compliance with each site’s terms and conditions. If an individual on the site has indicated they wish their information to be private and not mined, ASI respects the wish and does not mine their data. Privacy is preserved using two methods: data anonymity, k-anonymity and differential privacy." — From the briefer.
What is Polly doing in the election?
Alongside creating these representative samples that are helping paint a picture of what the outcome of the Ontario election looks like, as it evolves, part of what Polly has done is determine 14 fragile Progressive Conservative ridings that Andrea Horwath could potentially win over if she were to visit.
READ MORE: Making microwave technology much cheaper and better
Earlier this month the Liberals thought they had a bombshell when they claimed that the NDP election platform budget was off by about CAD$5.7 billion, but according to Kelly, Polly — from the data — has analyzed that only eight per cent of people thinking of voting NDP in the upcoming election are talking about that, but the news is not changing their mind right now.
"What the AI is seeing is that the more she gets out there...she gets more and more popular," said Kelly of what Polly has 'found'.
Where else is Polly?
Polly was working hard during the 2016 US Presidential election, the Brexit referendum and during the Scottish referendum in 2014.
Other projects include ASI using Polly to held predict suicide risk in Canada.
More about Ai, ml, machine learning, Artificial intelligence, Technology
 
Latest News
Top News