Why Social Media Analytics Cannot Rely On Machine Learning

A human touch is still necessary


Social media has been the main driver of public conversation for the past decade now, with Facebook having been founded in 2006 and the likes of Twitter and Instagram entering the fray in the ensuing years. According to estimates, in 2016, 78% of all Americans used social media - more than 200 million people. Worldwide, the number amounts to some 2 billion.

Social media is essentially the internet in its purest form, a technology that brings people from across the planet together. This probably seemed like a good idea at the time, although it failed to account for the fact that people in groups tend to be stupid. A person is smart. People in groups are dumb, panicky, dangerous animals. And this can be seen daily on newsfeeds on every platform, which have become a dystopian nightmare of hate, mistruth, narcissism and outright stupidity.

People in groups are also easily exploited, something that marketers and businesses recognized from the outset. Having discovered that social media was an extremely convenient way of advertising to a highly engaged mass audience, they set about monitoring and tracking variables to segment them based on a variety of metrics so that they could tailor and target their campaigns to more specific groups. The social media companies also made this easier, using algorithms to learn what kind of content people wanted to see so they could essentially curate their feeds, showing them little outside their existing likes. This has had a variety of negative consequences, entrenching people’s world view and restricting public discourse.

We are now at a stage where this is evolving even further, with AI and machine learning algorithms taking over from human marketers, and in some cases human users. This has strange implications. Go on Twitter now and you have bots liking tweets sent by other bots, retweeting tweets from other bots, an infinite Russian doll of bots engaging with bots to no apparent end. Twitter is a mess of automated tweets sent by people who haven’t even bothered to read them themselves.

This is not to say that AI is not making the job of social media easier for marketers. There are, undeniably, many areas where it is a benefit, particularly in terms of analysis. There are now thousands of channels across multiple devices, making it practically impossible to measure user engagement manually. Every minute, social media users submit over 347,000 updates to Twitter and ‘like’ more than 4 million things on Facebook. No one person could visit every single social media site and manually count likes, retweets, and shares, monitor comments for sentiment, and record various data from hundreds of other variables - and nor would they want to. That’s to say nothing of influencer analysis, trend monitoring, deep filtering, and data segmentation.

However, there are many problems with the technology in its current incarnation. While algorithmic analysis of the social media landscape has become more sophisticated, machine learning cannot yet be trusted to classify and categorise all the nuance inherent in human communication.

The human language is complicated, and it is even more complicated on social media - particularly Twitter as words and grammatical structure has to be adapted to fit into the 140 character restriction. Slang, new words, sarcasm, private jokes, and double meanings are all central to how we communicate with one another, and such nuances are still beyond the ability of computers to understand, making it impossible for them to properly sort, classify, and rate social media information for relevance and underlying sentiment. This causes meaning to be distorted and uncertainty around data sets that have to be accurate because they so often form the foundation of decision-making.

There are several recent examples of its failures - including one often held up as evidence of its success. Polling had a rough time of it in 2016, with notable failures to predict both Trump’s election and Brexit. Many pointed to social media analytics as the alternative. MogIA, for example, looked at 20 million data points around engagement on online platforms like Google, YouTube, and Twitter to predict a Trump victory. This was, however, a prediction that he would win the popular vote, which Trump actually lost by 3 million. It did so because it was unable to account for the high degree of negative engagement on social media, with many vocal in their dislike of the president. While it was, in the end, correct, it was only correct on a technicality.

Ultimately, AI is not going to render human social media analysts redundant any time soon. Marketers are still needed to audit sentiment for accuracy and to feed in data to help machines get better at building a comprehensive view of language so that it can analyzed. But language is such a fast moving target that it is likely people will never not be needed. Which leads to the question, have machines made social media better. In order to gain a nugget of value now, you have to sift through a pile of trash. This is, however, largely a pile of trash of the machine’s own making. At what point do they end up destroying social media, and what can be done to stop them? We now have cleverer machines trying to wade through a mess created by stupider machines to try and flog something to a human who may or may not even be there. This is the future, and it is madness.


comments powered byDisqus
Pins small

Read next:

EHRs: Linchpins For Big Data