getting the same sample sizes, we're still approaching the science and the same way, even though the population is changed a lot. now with social media, finally, we have the opportunity to involve this science for the 1st time in 50 years. i traditional pulling his truck out on several recent election predictions. we're going to win the white house. we're going to take it back. perhaps most famously in the 2016 us presidential election posters overwhelmingly predicted hillary clinton to be donald trump. if the lines are long tomorrow, please wait me trumps. when demonstrated that posters had missed the boat, they predicted the popular vote but missed the tight local races in swing states, which delivered trump and electoral college victory. clearly, their models were just completely off. the actual percentage of the votes predicted

by the polls is actually what happened. but we weren't measuring the electoral college. obviously, you go through, you know, a dark night of the soul. whenever these kinds of things happen. it's too expensive and inefficient to poll regularly by traditional methods and each electoral district known in canada as a writing. but social media is huge, scale makes that possible with a calling about politics and social media or all your opinions or social media is now kind of the norm that's become a normal way to communicate in the same way that people use. i think long conversations on the phone frequently to now, of course, back in the day you can just tap people phones you had to call up and ask them, but opinions but, but now the thing was social media. it's interesting is when people are communicating their opinions on social media, often they're doing it in a form that is public. and that allows us to see what they're saying and to then derive their, their opinions from that. but not all social media platform have data that are

relevant to white. we use a sample for data source, depending on the problem that we want to look at. we have found that for elections, we get the strongest signal from many posters and social scientists question the value of data from twitter platforms like quarter. and facebook is basically a group of self selected people. not every canadians are on cultural media. and even on social media, got all of them use it to discuss politics. so get it. and basically you're getting a slice of smaller life of the public. i think it's like 12 or 15 percent. use twitter for news 20 percent. use twitter overall in my own lay understanding of that. i tend to think about to lease journalists and like high engaged, super participants that are on twitter. so not everyone is on twitter. let's take the loan number, 17 percent. that is still a whole lot bigger than the current 9 percent or less. you get on diets. 17 percent

of a country of 36000000 is about 7000000 people that we have access to me. white things that a i can fit through the large but typically young and urban user base of twitter and canada and none the less make forecasts that are representative of the entire population. less than 3 weeks to go until election day. and i know at some point, you're artificial intelligence pulling system comes to a conclusion about what you think is going to happen on the 21st of october to early still, i presume it's too early. what both well for mr. trudeau is that the liberals have consistently started this campaign. been in 1st place, 1st place in the popular vote, or total vote, i guess, as we call bandanna or c, count speak count the different. okay, so we're doing each individual writing so and that's as we saw on the ontario election, it makes a big difference just because you're leading in the total vote count in the polls.

doesn't mean you're going to win exactly the vote efficiency. yes, it was with whereas that in a 1st post test proposed system, that's right. talk to most people that work with data from lots of people. they will say i want to collect everything. everything everyone is saying on facebook and twitter and read it and look at it all. most critics assume air sy withdrawn what is called the twitter firehouse. if so, why we need a way to filter the huge volume of real time data and collect only the data useful to predict elections. given the magnitude of the challenge. critics dispute whites optimistic claims. now if i all, i did, capitan that fire pull in that stream of data, you're going to get a very skewed distorted view. that's why we don't count. treat people

are big idea. is what the posters have been doing. works actually secret sauce. it make really, really good sam, the people in other words, rather than trying to clean up the twitter fire hose, white claims that he has found a way to bypass it before the holiday collects any actual data. poly 1st builds a representative sample. just like posters. those who say that you can look at social media and get to representation and what's happening at the writing level. it's an interesting concept. we're certainly exploring that ourselves. what i bet the farm to morrow and being able to do an election in which i predict every single right incorrectly based on social media. no. the 1st of all, prison people back to the writings difficult. it's difficult to figure out where is

that account tweeting from that old joke the, on the internet. nobody knows your dog well on twitter. nobody knows whether you are in china, russia, or right here in toronto. it's very different. i'd say yes for sure. that accounts come canada, a representative sample of canadians must consist of actual canadians. white thinks he can confirm his twitter users are legitimate. we use the information that people put on their account. so we see that they're in canada. great. they say, i mean channels. so we verify that the 1st test is your friends, also in canada. look at your history. look at what you're talking about. if you're canadian, probably talking a little bit about maybe hockey or other canadian thanks. if you're talking about the cold weather for posters, a sample is representative and random only when every member has an equal chance of appearing in the sample. imagine that you had

a whole big bowl of jelly beans and there was a red ones and blue ones and green ones and yellow ones. you don't need to count every single jelly being to figure out if there's red ones or whatever the colors were. i said, if you, if all of them are, are equally distributed across the entire big ball of jelly beans. if you stick in your hand and you pull it out and you put down those jelly beans, you should have some representation of what's inside the jar. if you do it 10 times, you'll have a pretty good representation of what's in that you're not because the jars getting smaller, but because the samples become consistent. so that's what sampling is really all about. the historically, poster sampled responded by randomly selecting names from a phone book. in the digital world. random and representative sampling requires a different paradigm. is a game called 6 degrees of kevin bacon and says,

i'll pick a movie actor and try to link them back to kevin bacon 6. and you can play the same game on facebook and twitter can pick any person any celebrity, and tracks link yourself back to them, and on average, about 23, and 4 connections. anyone in the world? holly's algorithm uses a mathematical computation to guarantee an average minimum of hawks between the twitter users that collects. that's how it defines virtual distance and randomness between users. the critics argue that still doesn't address the lack of diversity among twitter users. i can account for that bias and do it by saying this is what the census says, how my distribution should be. and then the algorithm, you say it differs too much from what you expect and you have to wait until balances to for example, when i see that my sample starting

a little bit lopsidedly younger, now just stop collecting younger now you pause know, add any more into your sample to you start to get older rural female. so instead of looking like the majority of twitter, poly sample is limited to a mix of users that match the actual demographics of each writing. why claims this addresses? what critics see is a tragic flaw in polly's main source of data. and that is the key algorithm behind poly. if ever a bombshell hit a canadian election campaign, it happened this week. photos of justin trudeau in black and brown face swirled a very complicated broo of opinion in the country. aaron kelly, what is poly telling you about the fall from all of this? it was a big bomb show. so poly saw mr to those numbers, the number of states feet count to fall by 25 feet yesterday. he is rebounding this

morning. he's up 5 so soon that of 20 feet losses, but it knocked him from majority territory into minority territory. jag meet saying the n d p leader was the biggest winner yesterday. we saw his, his poll numbers go up. so he has really, he really shot through this. so how does an algorithm go from tweet to projecting a drop in feet? what is a tweet actually mean? what does it mean when we're tweeting? we know people are making jokes in error there i, ron, people are showing more images and that creates a challenge. it's harder to interpret means, which is these shorthand graphics are often pop culture references. it's like, i don't know what's going on here and then i'm like, well, i don't understand this and my computational power is a living, breathing human. what is poor computer, you know, he's going to have to make sense of this. according to white. poly is not trying to interpret the meaning of tweets at all. so my 1st few for a's and 2 elections

forecasting. i ran the experiments and consistently what i found was just the plain old mentions out performed the positive or negative sentiment every time. ready all the tracks, the percentage of users in the sample who mentioned each federal party. then the a i compares those results to historical opinion patterns. after a scanned, broke, involving an old photo of the prime minister in blackface, mentioned to the liberal party spiked in relation to the others. without understanding the meaning of these tweets, polly interpreted changes and mentions a drop in projected seats for the liberals. i think paws are interesting because they're a flash point in the past, right? they don't predict the future there a moment in time in which you're getting a flavor of what the public is sort of thinking about the a i collect data constantly that makes it possible to identify patterns and user activity over time compared to poles that are held less frequently,

one of the biggest, most difficult things in pulling when you're doing these tracking calls is you see something happen for good or bad. they say, well it's one piece of data is going to do, then you have to wait another day. is that real or was it just a blip here? always sort of in that in that feeling of how many points makes a trend, is it one? probably not. is it 2? maybe 3. yeah. for for sure. but you know, where is that point? in order to identify a trend poly use as public polls as a reference point? she analyzes patterns of user engagement. whenever her selected twitter users mentioned political parties. then she compares engagement with each party standing in the pulse. based on those relationships poly predicts of her data indicates a trend is stable or not people. sometimes they'll be angry about something that an issue and they might be angry about it in a moment of time. the poly is able, in a very unbiased way to see through that anger and say,

how long will it take the population to get over this anchor and this hurt. and will it carry long enough to election day poly tracks, twitter engagement constantly, that continuous stream of activity is referred to as longitudinal data by studying data. and this way poly can identify patterns over time and then compare them to present activity. it also means poly can look backwards in time at the data to reexamine unexplored trends for the brown case, black scandal. i don't think there is a single poster that asked a question if you saw a party leader in a questionable costume in their twenty's. pretty change your opinion. i don't think anyone thought to ask that question to the wonderful thing about poly in the launch 2 studies is that i can go back, i can get an accurate measurement of exactly how you thought a week ago or

a month ago. and that's something you can't do a week later or 2 weeks later by asking a question. i would suggest that perhaps canadians don't really care about blackberries as much as they perhaps profess do. on social media, there was a lot of outrage and i would even call it performative outrage on social media in particular, but being like, oh my god, i can't believe this as i really can't believe those. you can't believe that, like i'm or, you know, a rich white guy from quebec would have dawned black face at a receipt private school that had a party themed and ruby and nights. like, i mean, the party in and of itself without the blackface was racist in the given how massive that story was when it broke a week and a half ago. it has virtually, i don't want to think disappeared from the media,

but you see so few references to it. now, is this story essentially over to the best that you can tell through your polling? i think to keep bringing it up for the other parties. keep bringing it up would be beating a dead horse at this point. i think he is recovering. and i think the slower recovery might be due to you know, a mix of issues and not just blackface chug me thing really impressed. a lot of people with his initial statements after the black face scandal broke, that goodwill is done now and as in gone, well, he's up to seats from before blackface, but that's down from having been up 9 before. so it's starting to, peter owed me like artificial intelligence and machine learning or so i use in different ways to mean different sorta magical things about computers. i think the way humans do modern meaning pretty much is just these very sophisticated and very flexible algorithm which allow us to process very large amounts of data with, with very complicated relationships between them and have very flexible models

which will allow the computer to sort of figure out, unlike the reason why or why not just statistics or collection of coke, the difference is that she's able to do this mostly on her own. i don't have to go in and write a rules and say now poly, when you see this, do this. she's able to figure things out on her own. polly's ability to learn is the basis for how she predicts future voting outcomes. ah, the key to machine learning is a process known as training a statistical model. talk about training a model. we mean, we have some data where we already know the answer. so maybe as how people were in the last election, or maybe you know, how many people did or didn't buy guns, but we already know the answer. and then we use that to drive in france as a okay. it seems like man who live here are more likely to vote for that person or

many subscribed to this magazine are more likely to buy a gun. and we learn those relationships. and we say, let's see if we can use our model that was trained on a previous data to try to predict what's gonna happen next. ready election night 2019. ready almost $18000000.00 canadians have cast their ballots, including the advanced pulse that represents a turn out of 66 percent, several percent higher than expected. the rental form the government, but the conservatives win the popular vote. days later, it's time to compare the actual results with the seat predictions federal election, 2019 wrapped up of course this week with a liberal minority government. this is the national seat projection that you gave us would show the liberals in arranged between 116 and 173. but essentially at 145, the tories at 123. the n d p at 48 the block at 46,

and the green somewhere between one and 5, and you have the possibility there of a couple of independence and a people's party. now let's flip that over. and this is what the national seat projection was on sunday. so the last one was saturday. this one is sunday, the day before the election. and you could see that things have changed. you've got the liberals now at $155.00 tories that $118.00. and then let's roll them monday. and this is what the actual election results were. 157 liberal seats, 121 conservatives. 32 block, 24 new democrats, 3 greens, one, jody wilson, re bold and nothing for the people's party. if you ever want to evaluate how well someone did predicting election, i think the fair thing to do is to compare them to other people who try to predict the election, yvonne company and better than all the other companies. and that's probably a pretty good result in the day before the election. some traditional posters were only predicting the popular vote. others were also willing to predict

a liberal minority, but few publicly predicted a narrow range of seats. ah, according to pollster 338 canada. they correctly predicted the outcome of 299 of the 338 writings according to advanced symbolic poly correctly predicted 308. was this quote unquote, an easy election to predict. it wasn't an easy election predict, but it was predictable, right? and i find it, you know, it's interesting because everybody was saying was so close, that's the, that's why being able to look at the seat count and do the writings is it's becoming almost an necessity if you want to do pulling. we saw it at 1st with the trump election in the states, and that's been consistent. it seems that elections are getting closer and closer. and so obviously when it's a minority government, it's a little bit harder to call. but this was consistent for poly that she always

thought as a liberal when, if you look at her election results, we did really well on the 2019 election. but there were some writings. we got off. it comes down to we're making assumptions that all parties are campaigning equally on the last weekend before the election in all writings where they might have a chance. all the political scientists out there are probably laughing at me right now. turns out that's not true. turns out political parties tick, writings that strategically they want to win. and we haven't taken any of that into account. it's why i would love to have a real open dialogue with the people that know the political science. both white and kelly probably claimed that poly spinal c prediction was more accurate than most established posters. although white does admit a lack of expertise and how campaigns work to critics of an a i method this

admission reveal some naive and damaging political assumptions. but to supporters, holly's relative accuracy and projecting seats shows great promise for whenever i hear people say i've got the things i sit back and say ok. * not the thing well, you know, the big publicly traded research corporations that have more money to do this kind of stuff than anybody else has to do it. not to mention the google's in the, in the facebook's, in the twitters and all these other people who have all this money and all the best scientists all working on this kind of thing. they don't, they can't predict based on their platform, better and posters in status tissue still demand more proof. they continue to doubt the value of twitter as a primary data source. and they insist that poly established a much longer track record of predicting elections before they concede this method

is reputable and accurate. so the common saying this uses a defamation, people buy products and so on. that says past performance is not an indicator of future performance. and i think that applies to predictive algorithms as well. if they've done well in a last couple of elections, i think that works in their favor, but still not a guarantee. and as they do better and better in more and more elections, i think you can start to trust them more and more according to a s i. polly's prediction model will only get better with more election data. and with input from social scientists to refine their algorithms, we've come a long way to get to where is that today and a lot of what has helped a i get to where it is that today is the internet, the explosion of what the explosion of public data comes with risks. social media is free exchange for users, privacy. personal data is then available for marketing. critics also warning that

platform, they're using the data against their own users by making the platforms more addictive. when the car came out, there was no seat belt. there was no airbags, but when we discovered that, oh my gosh, these cars and getting the car accident, they can run people over. we are demanding better breaks. we start demanding airbags, we start demanding seat belts. the 1st 1015 years of social media, the argument was, let's not put handcuffs on this company because we want to see where and what they can do and how to help society. but we're now discovering all the car accidents. we're discovering all the harms that they can do. so for me and i think the next few years is going to be debating what does regulation especially look like. i think the big problem is we're thinking of campaigning in an analog world. when in fact, it's all automated and digital, and neither our language nor our laws really anticipate the power that facebook has in the way in which individuals are political parties or foreign governments can

use that power to tip the scale of the election. if not just money, people's perception of the election, because the whole point we have an election is to legitimize the if people believe rightly or wrongly, that algorithms have taken over the election. well, then they're not going to see the results as legitimate. they're not going to see the government as legitimate, therefore it's going to make government's ability to actually govern that much more difficult in many ways, using data to get people to vote and pulling people to participate in politics is fairly beneficial. in fact, if you were to ask me, what could data be good for? it's like getting more engagement in politics. and so i don't think that by definition, political advertising is a problem where i think is that the way that online advertising works in the fact that requires so much surveillance to develop ad profiles. and that the way you're doing targeting so much can fraud. those are the kind of relations i think

canadians are feeling creeped out about the way in which technology and surveillance is playing a growing role in their society. but i think they feel powerless. so we've had this notion of, if you want privacy, stay at home. if you want privacy only tell, but if you decide to participate in the public square, you put information out there. you will be held to have said it's okay for anybody to do anything with that information. and that is what just does not work in the modern world. i think it's really quite pernicious to think that we'll, we'll solve this with consent and we're just going to change the consent in or consent out the fly. and that is that still the individual against this massive organization. and the great divide is they know a lot about you. you don't know a lot about them. we invented all of these structures. and we can invent something else. polly's election predictions depend on the ongoing popularity of twitter and

easy access to its data, but at new regulations to protect user privacy or restrict the social media business model came into force. they could profoundly change the way people use social media. i don't believe social media is going to become more private. there will be a need for public networks where people can go create content that anyone can view and anyone can engage with kind of making a big debt on it to me. i don't know. i think maybe the traditional pulses are afraid of because it's the future. there was a twitter shot that microsoft name in the chat box, but it was a demonstration of language. and it was a really, really sophisticated,

wonderful lot of people got wind of this chap on twitter county. we started saying, what can we do to teach this a i to be the most offensive repugnant twitter account on the planet. and they succeeded. little a, i became a horribly racist xena phobic microsoft had showed us why do we as people feel the need to be miserable and mean bad to art and show that example to the machine? ah, when you see bad children usually so when it comes to a i, when it comes to all of this, to me, north star, i have to be a good parent. have to be the best parent for polly's me.

i oh i i the, we don't want to see them, but they are there in our streets. water. even now we're here. unseen about the threats you're facing. the heroes taking the standards

on d. w. africa. the great green law. the help when finished, it will be more than 15 kilometers wide, a 1000 kilometers long, a lush green thread to combat deserted acacia. what's been accomplished since 2007? how can the initiative be improved? we talk to the projects director infiniti. i don't go for a minute. ah, was the little guy that is a 77 percent the flat for african issues and share ideas.

you know, we are not afraid to detach from africa. relation is ruling class. young people clearly have the solution. the future is 77 percent. now every weekend on dw the w news. these are top stories. organizers of the european football championships have refused in application by the munich city council. to have the stadium, they are eliminated in rainbow colors for germany's game against hungry sports governing body said no because of what it called, the quote political context. the gesture was meant to protest the new hungarian law

widely regarded as discriminatory against l. g. b, t. people.