I’ve been thinking about Facebook’s new dating app. I suspect that it has the potential to be very good—or very, very bad.
Facebook is a big data company: they make their money because they can very precisely figure out what users will respond to. What if they applied that to online dating? Maybe it will look more like other dating apps, but remember how much Facebook knows about people. In particular, at this point it has many years of data not just on individuals, but on which of its users have partnered with which others, and (to some extent) on how long these partnerships last. That is, rather than code an algorithm that effectively says, "you two match on the following N points on your questions and answers", Facebook can run a machine learning algorithm that says "you two cluster with these other pairs who went on to serious relationships." (Three times already when typing this, my fingers typed "dataing" instead of "dating". Damn, make that four!)
So what’s wrong? Isn’t that a goal of a dating app? Well, maybe. The thing about optimization is that you have to be very careful what you ask for—because you may get exactly that, rather than what you actually wanted. What will Facebook’s metric for success be? A couple that seriously pairs off, e.g., moves in together and/or marries, fairly soon? A couple that starts more slowly but the relationship lasts longer? A bimodal distributon of quick flameouts and long-term relationships? (Facebook says they’re not trying for hookups, so I guess they don’t need to buy data from Uber.)
There are, of course, all of the usual issues of preexisting human biases being amplified by ML algorithms, to say nothing of the many privacy issues here. I think, though, that the metric here is less obvious and more important. What is Facebook trying to maximize? And how will they profit from the answers?