Facebook Knows When You're Sad (#2)

#102 – Facebook Knows When You’re Sad

I knew what Facebook was doing.  As soon as it gave us the option to put a standardized set of a emotions next to our statuses, I knew crafty ad men were at work.  But first, let me back up.

Computers (and those programming them) have been working hard for decades to better understand the quirks and nuances of language.  Natural language processing (NLP) is, depending on who you ask, another step towards the robot apocalypse, an exciting path towards “talking” to our devices, a new method of documenting and understanding our feelings, or a chance to make bank. 

But, regardless of the implications, NLP has a long way to go.  A computer program could very confidently determine a melancholy tone if you type “I am sad”, but what about “I am crushed by bowling balls that represent the weight of my problems” or “Being on hold for hours with Comcast is SO MUCH FUN”?

I’m going to avoid the additional hurdles of spoken word processing (like accents and inflection) and instead stick to text.  The example that got me thinking about this was Eric Gilbert and Karrie Karahalios’ study on LiveJournal blog entries and their ability to predict the stock market (which I read about in Predictive Analytics – great book and easy read!).  Eric and Karrie did not attempt to guess the emotions conveyed by certain words, and instead compared million of entries with Livejournal’s option to label posts as “happy”, “sad”, “anxious”, and more.  Finding correlations between, for example, the use of phrases “really scared” or “care-free” and the likelihood of placing a positive label, allowed the two men to predict the anxiety level of a post by analyzing the text, even without a labeled emotion.  The analysts then aggregated recent posts on LiveJournal, and later combined other blog sites, to form an overall anxiety index, on the hypothesis that higher anxiety among internet users would predict stock market drops (and vice versa).  In case you were wondering, this approach was successful for a while, until it wasn’t. Now, it appears Facebook wants to build its own emotion-prediction model for the purposes of ad targeting.

How deeply will Facebook’s system penetrate our heads? It’s hard to say, but the most valuable (and coolest) part about predictive models is their potential to find unintuitive, even bizarre relationships.  Maybe a model will find something totally unexpected, like if people click on more political ads when hungry, or buy more pens, but fewer staplers when feeling excited.  It would be unlikely a human, lurking on our statuses, could make these connections, even if they worked as blisteringly fast as a computer running a million little optimization experiments. The social network could then optimize promotions for its partners by placing them only in feeds when Facebook users are in the mood to see them.  This is a ridiculous level of targeting.

Privacy-zealots, big data-skeptics and some casual internet browsers will find this inherently creepy and even dangerous.  This is a topic for another post.  However, I’m not too bothered, and am inclined to let Facebook try and guess what I’m thinking, and give me ads I’m slightly less likely to ignore.  Even so, those sneaky stat guys might find I’m privy to their game, and am happy to mess with them.