Jean Christophe Dornstetter (Marie Quantier): “AI makes it possible to survey more people”

by bold-lichterman

On November 8, Donald Trump was elected President of the United States. Big surprise, all the polls announced Hillary Clinton as the winner. Are polls no longer reliable? Can artificial intelligence help to be more representative of the opinion of respondents?

More details with Jean-Christophe Dornstetter, Head Of Artificial Intelligence at Marie Quantier.

FrenchWeb: What can artificial intelligence bring to traditional polling techniques?

jean-christophe-dornstetterJean-Christophe Dornstetter, Head Of Artificial Intelligence at Marie Quantier: Before answering this question, we must understand what polls are today. The objective of a survey is to determine – or at least to estimate – the state of opinion on a given question. The ideal method would be to interview everyone involved, but that would be very expensive. This is why polling institutes have recourse to statistical mathematics to estimate the state of general opinion from a small number of individuals polled. Thus, many polls that appear in the press are carried out on samples of around 1,000 people.

The difficulty of the surveys is the representativeness of the chosen sample. In addition to counting the response of a respondent, it is essential to know his profile. The institutes first define a set of characteristics (age, sex, socio-professional category, city / suburban / rural, etc.) of which they know the representativeness within the French population (via the census, figures from INSEE , the results of previous elections, etc.) and question them in different ways (at the end of the ballot box, by telephone, by internet, etc.). The targeting of the people questioned, at the heart of the method, can be refined today thanks to artificial intelligence. It’s actually a two-part transformation – as so often when it comes to artificial intelligence.

First, the arrival of big-data, in other words the possibility of collecting a lot of data compared to traditional techniques. This is made possible with the daily use of the internet and mobile applications, which are the site of a battle – admittedly silent but fierce – to retrieve as much information as possible from users. And yes, Big brother is watching you. The collection of personal data is central in business models giants of the Web. These new data (articles read, videos viewed, podcast programs, websites visited, etc.) enrich the data collected during traditional surveys.

Processing these mountains of data is the second component of the transformation. What we call “artificial intelligence” is the design of very sophisticated algorithms that allow navigation in this immense ocean of information and derive statistics and correlations.

With AI, will the methods change?

Basically, the way you do a survey will be the same. Artificial intelligence will be especially useful upstream to improve targeting (ie improve the representativeness of the sample of individuals surveyed), and downstream to extract and consolidate data.

Is this the end of surveys for the benefit of shared data?

With the increasing digitization of the French population, the representativeness of Internet and mobile application users (from which we will be able to extract data) will strengthen. Thus, the extrapolation of the chosen sample to the overall population will be better. We can also survey more people.

The diversity of data types also allows us to be imaginative in bypassing traditional survey methods. Some surveys can be corroborated simply by measured behaviors. To take a recent example, the strong participation in the primary from the right and the center was predicted by the large number of visits to the website to find his polling station.

The answers given during a survey are based on declarations. There may be a bias depending on the questions (for example: the far-right vote is under-reported in opinion polls). The data collected is more objective (it is tedious to rig your behavior on a daily basis), which gives them a qualitative advantage.

Polls and shared data go hand in hand. I believe in a coexistence / synergy of the two types of methods: surveys that can help to know what to look for precisely in these mountains of data; and this data can also help determine the subject of the survey.

Read also: