Institute For The Future

Take This Medicine, Courtesy of Your Social Network

A great feature in The Economist highlights the variety of ways businesses and researchers are looking at analyzing the intricacies of our social networks and digital trails to understand who influences us, who we influence, and what that could mean for the world at large. This isn't a new field, per se, but the breadth and subtlety of the analysis, as well as the potential quality of their conclusion, is pretty mind-blowing.

 

Take this example, which suggests that police officers can prevent crime by monitoring Facebook and Twitter.

 

The police department of Richmond, Virginia, has pioneered the use of network-analysis software to predict crimes. Police officers know that crime increases at certain times, such as on paydays and when there is a full moon. But the software lets them analyse the social networks around suspects, such as dealings with employers, collection agencies and the Department of Motor Vehicles. The goal, according to Stephen Hollifield, the department’s technology chief, is to “pull together a complete picture” of suspects and their social circle.

 

Party plans turn out to be a particularly useful part of this picture. Richmond’s police have started monitoring Facebook, MySpace and Twitter messages to determine where the rowdiest festivities will be. On big party nights, the department now saves about $15,000 on overtime pay, because officers are deployed to areas that the software deems ripe for criminal activity.

 

 

Or, for that matter, consider a different example from the Economist, which focuses on how companies can identify, and tailor pitches to more influential customers:

 

TELECOMS operators naturally prize mobile-phone subscribers who spend a lot, but some thriftier customers, it turns out, are actually more valuable. Known as “influencers”, these subscribers frequently persuade their friends, family and colleagues to follow them when they switch to a rival operator. The trick, then, is to identify such trendsetting subscribers and keep them on board with special discounts and promotions. People at the top of the office or social pecking order often receive quick callbacks, do not worry about calling other people late at night and tend to get more calls at times when social events are most often organised, such as Friday afternoons. Influential customers also reveal their clout by making long calls, while the calls they receive are generally short.

 

 

In part, what struck me about these examples was the sophistication and granularity of the analysis--the major insight doesn't come from looking at which people are connected to each other, but by figuring out who gets to have long phone conversations whenever they want them. Dictating the time and length of a phone call--that's a sign of social influence that a business needs to cater to.

 

In this sense, the ability to distill profound conclusions from a mess of data reminds me of this semi-recent article about the ability to detect depression via blog post. And not, you know, by a trained psychotherapist reading a series of blog posts, but through an incredibly sophisticated algorithm:

 

"The software program was designed to find depressive content hidden in language that did not mention the obvious terms like "depression" or suicide," explains Prof. Neuman. "A psychologist knows how to spot various emotional states through intuition. Here, we have a program that does this methodically through the innovative use of 'web intelligence.'"

 

For example, the program spots words that express various emotions, like colors that the writer employs to metaphorically describe certain situations. Words like "black" combined with other terms that describe symptoms of depression, such as sleep deprivation or loneliness, will be recognized by the software as "depressive" texts.

 

 

It's enough to imagine the sort of future where a pharmaceutical company's algorithms can read through your Outlook calendar, notice no one accepts your meetings, sees your Facebook status updates seem to indicate a level of frustration, and sends you an offer for a free sample of the latest anti-depressant.

 

But I think there's a bigger danger here then the idea--which, again, isn't new, and which I remain skeptical of--that intelligent algorithms can predict our futures. And that danger, I think, is of the self-fulfilling prophecy.

 

Take, for example, another use of social data analytics highlighted by the Economist: Using social data analysis to tweak things like credit scores and decisions about whether or not someone might qualify for a loan. It might make good statistical sense to look at someone's social network and deny them a loan, but the end result is to make it difficult for someone who might very well be a decent credit risk to get a loan--which would, in turn, probably lead to a series of depressed blog posts and an offer for a free month of anti-depressants.

 

Indeed, the chance that small signals--like an social circle that isn't credit worthy--can snowball into other problems over time strikes me as a huge risk of this sort of business strategy (albeit one not borne directly by the business.) And even if the analysis is solid statistically, it doesn't make it a good analysis at the individual level--since, almost by definition, if you're trying to analyze tens of millions of people, you're bound to wind up with some weird, and potentially very problematic, outliers.

 

I also think it's worth echoing a point by computer science grad student and science writer Greg Fish:

 

But the problem is that a very large data set about consumer behavior really only tells you what consumers like at the moment and the emerging fads of the day rather than alert you to what’s going to be really popular and marketable in the next six months to year, giving you enough lead time to develop and test your product, as well as its marketing. The idea of looking for statistical patterns in complex data sets has been tried before on the stock market with very mixed results. Pretty much all systems that billed themselves as excellent predictors of where the market will move tomorrow, or that week, have failed....

 

Ok, the stock market is one thing. Why couldn’t we use a stream of consumer data to make predictions? The prescient wonk approach to data analysis assumes that humans are more or less rational, and what they do and say now, can be extrapolated into the near future. However, we’re far more messy than that, and what we say in public isn’t always what we do in private. No data mining is going to explain why the very same people who post a long winded rant on their personal blogs about the demise of good literature own, and love, every single book of the Twilight series. Or why so many mediocre, widely panned creative works gain the success they do. All you’ll see are the double standards and contradictions writ large across your data set, tampering with all your significance tests. In effect, you would be trying to predict the actions of people who change their minds day in, day out, quickly embrace and abandon trends and fads based solely on how they feel over the last several months, indulge in guilty pleasures, and jump on bandwagons depending on how close they are with certain friends, whose relationships can change at any moment.

 

 

None of which is to doubt the substantial value in understanding social influence and network dynamics, which strikes me as the first real opportunity to quantify and operationalize the very real, but frustratingly slippery social sphere. At the same time, our social networks--or, for that matter, anything we enter into a computer--are just variables to be analyzed. While more statistical analysis is likely to lead to better understanding, we should remember that this sort of analysis is going to be speculative subject to error.