In February 2016 Facebook extended its famous Like button to include emotions: love, haha, wow, sad and angry. For the first time users could not only express their pleasure at reading a Facebook post, but their displeasure as well, and with shades of grey. Facebook also opened up an API to make this data available to developers.
At AI Mining we have looked at and analysed this data extensively. We focused on news data as it tends to show a wider range of reactions. For example, fans of musician Taylor Swift tend to love everything she posts, but when they react to a CNN news post, their reaction says more about the content of the news headline than about their feeling for CNN itself.
As part of our research, we ran the news data through various statistical correlation algorithms, to see if there was a significant relationship between the different user reactions. These reactions include not only the five emotions but also comments, shares and likes. This article discusses some of the statistically significant correlations we found. Remember this analysis applies to news data only.
First we’ll show you the results of our research graphically, because it looks great. Then we’ll point out some of the surprises and show the raw numbers. The graph below links the 5 reactions and 3 actions to each other according to strength of correlation:
Likes and Love
The strongest correlation above is the thick red line between the like button and the emotion love. The strength is 0.87 out of a maximum of 1 (all the numbers are at the bottom of the article). This particular correlation isn’t at all surprising but we sort of have to mention it.
Comments and Haha, Wow and Angry
Commenting is roughly equally correlated with the reactions haha, wow and angry (0.72-0.74) but much less with love (0.64) and sad (0.59). So if you want your news to be commented on, make it funny, surprising or riling. This is not so surprising – the more provocative the news, the more comments it will gather. Simply making people feel good/love or bad/sad does not drive user engagement.
Shares and Haha
More interesting is that the reaction mostly strongly correlated with shares is haha. In other words, funny news is more likely to be shared (0.77) than surprising (0.51) or angry news (0.52).
Sad and Angry
Between the user emotions, unsurprisingly the highest correlation is between sad and angry (0.66).
Haha Is Ambiguous
More unexpected is that haha also correlates fairly highly to sad (0.61). This is much higher than its correlation with love (0.52). In fact our analysis shows that love and sad/angry are the only very lowly correlated emotions (0.25).
This has caused us to rethink Facebook’s emotions. We initially thought that they purposefully chose two positive emotions (love and haha) a neutral one (wow) and two negative ones (sad and angry), in an attempt to balance out the emotions. To put this in context, psychologists have long thought that there are six or seven basic human emotions, and they are mostly negative. For example Paul Ekman’s seminal 1984 study posits: happiness, surprise, anger, fear, disgust, sadness – arguably 1 positive, 1 neutral and 4 negative.
Our analysis of Facebook emotions shows that love is the only definitively positive emotion, sad and angry are negative, but haha and wow are ambiguous. Maybe Facebook did intend the emotions to be balanced originally, and this has taken them by surprise as well. In any case, it fairly easy to explain: people often laugh or express surprise at others’ misfortune.
It is worth mentioning again that this only applies to news data, because when reading a CNN headline, the reader is not personally connected to the subject of the news. If it was their best friend, perhaps they would be less likely to laugh when something horrible happens.
Actions to Actions
The analysis above focused on correlations between the 5 emotions and 3 actions, and within the 5 emotions. Within the actions, comments, shares and likes are all well correlated to each other (0.74-0.79). This just means that articles which receive lots of comments, also get lots of shares and likes.
The table below shows all the correlations numerically. We used Spearman’s rank correlation coefficient (or rho) for this non-parametric analysis. This was based on 41,000 news headlines downloaded in mid 2016. Reactions and actions were normalised within each headline before correlating.
In summary, if you want your headline or ad to be clicked and shared, then try to make your readers laugh or make them shocked or angry. Even better, use our technology to do it.