I am not an engineer
Must read:Is there a real future in data analysis for self-learners without a math degree?

Social media analysis in the wake of the French presidential elections

Last year we woke up at least twice only to discover the voting results that we least expected. Have social media failed us or have we failed social media? What has changed since the last elections? Yesterday night, a panel of polling, social data and political experts discussed the state of social affairs at the Social Media Club in Paris. And no, Filteris was not mentioned during the debate.

To fully understand the challenges industry leaders are facing in election forecast nowadays, let us first have a look back at how things were handled back in the days.

Before polling, there was electoral geography

Before the introduction of polling methods in France in the 1960s, researchers would analyze election outcomes by linking them to socio-economic factors within a territorial unit. It was not about people and profiles, it was about social parameters prevailing in a certain territory. Still today, when presidential candidates visit a city, neighbourhood, small town or attend a public event on their campaign trail, they obviously take into account the local socio-economic parameters in their speech.

A famous theory by André Siegfried included even geological factors according to which areas with granite stone layers vote right whereas areas based on chalk tend to vote left. As a matter of fact, villages on granite soil were more scattered, thus favoring rurality, big land-ownership and the central role of local dignitaries. Villages on chalk soil however, developped as more densely populated settlements. Modern-day electoral geography would include underlying elements of the economic, cultural and social structure such as connections to other territories, etc.

The three polling revolutions

Polling methods were introduced in the 1960s in the context of a modernizing political landscape. The 1965 elections opposed different political generations: political giant and war hero Charles de Gaulle and newcomers such as the later center-left president François Mitterrand and Jean Lecanuet, who tried to import US political culture into France rather unsuccesfully.

Polling shifted the focus from geography to the voters as individuals. Socio-economic factors became of paramount importance, which led to a more systematic profiling process of the voting population.
In the beginning, Polls were conducted in face-to-face interviews with the participants. The development of telecommunications triggered the first technical revolution as pollsters started to carry out their interviews on the phone without physically meeting the participants. This increased the productivity of pollsters dramatically, thus reducing the costs of public opinion research.

The second revolution came with the rise of the Internet and online polling, which introduced scalable non-live polling and broadened the potential for refined profiling and individual behaviour tracking. Every method introduced a particular form of bias, which is one of the reasons all three polling methods are still in use today.

The third revolution would bring another method which does not even require anymore to engage into a conversation with the participant as it merely consists of listening to online conversations. Whether social listening is an extension of polling or should be considered as the beginning of a completely different field of methodology is yet another question.

Today’s challenges in social listening

Guilhem Fouetillou, co-founder of the social listening company Linkfluence, told a stunning story about how listening to the “social web” before the French presidential elections of 2007 actually led to accurate predictions.

“Our predictions based on social listening were more accurate back in 2007”

His team started in December 2005 to build a method allowing to follow public opinion and measure the popularity of the presidential candidates. Today’s leading social networks were not yet around in France at that time and social interactions happened mostly on blogs, forums and in the comment section (‘member?). Their algorithms were basically about deducting popularity by linking vocabulary to their tone and frequency. This method worked in 2007 but has not worked ever since. But why?

Fouetillou believes that bloggers and forums were not instrumentalized by the political parties back then .The problem of translating social conversations into a measure of popularity starts when the conversation or its channel have become a communication tool. Communicators have followed the path of Lazarsfeld all the way down to the online public space and even beyond.

This might be why pollsters and social listeners now cautiously present the “most active” and “most followed” candidates rather than the “most popular ones”. How to automatically distinguish a “natural message” expressing genuine voter approval from an “sponsored message”, which does not imply any approval at all?

One of the basic methods consist of choosing spontaneous messages only, such as the ones published during events. According to Fouetillou, live tweets are more interesting in this perspective as they are a spontaneous act of communication. However, the never-ending horse race continues as communicators are training political activists on live tweeting to inflate the “social” presence of their respective parties and candidates.

Pollsters and social listeners, why so discreet?

The cautiousness of the debaters was simply striking. One of them kept reminding the audience that his company was not into predictions, even though it has been publishing barometers on voter intentions. He also reminded that polls were mostly right about the US elections, but predictions were not. The complex voting system in the US might have contributed to the wrong predictions.

In France, on the other hand, polls have been showing mostly adequate figures in recent elections (European, municipal and regional elections). In addition, the voting system of the French presidential election is less unpredictable than the winner-take-all system in England or the electoral vote system in the US.
In the end, the whole panel agreed on not making any prediction that night.

Another implicitly black-listed topic was turn-out prediction, which turned out to be a deciding factor of Trump’s election. The researcher Thierry Vedel was quick to remind that the act of voting is a combination of having an opinion and of expressing it through a certain behaviour. As such, it is influenced by a series of various factors. For instance, the second round of the French elections takes place on a long weekend, during which the more privileged voters are more likely to leave their voting residence und therefore abstain from voting for their favorite candidate.

Of course, one can always compare past vote intentions with actual votes or measure the overall engagement level among voters by analysing their conversations in association with vocabulary indicating either vote abstention or participation. But, there again, the panel remained surprisingly discreet about how social data is currently used for turn-out prediction. Is this a regulatory matter?

Their secrecy might indeed be fuelled by the restrictive regulation provided by the law on “IT and liberty”. Whereas Vedel underlined the importance of such a regulatory framework in today’s world, industry leaders could barely hide their frustration over the “unused potential of social data”. As a matter of fact, listening to social conversation 24/7 has never been easier and cheaper. Linkfluence for instance have saved about 40m messages linked to the upcoming presidential elections, providing exhaustive data for in-depth analysis. Individual profiling of voters and tracking of their behaviour over time would reveal more of the underlying trends and tendencies, which are still extremely hard to grasp.

Leave a Reply

Your email address will not be published. Required fields are marked *