Thinking a bit about using machine learning for insight into human decisions

2017-6-30 01:32:01

This post was inspired by (at least) a couple of things:

  1. Listening to danah boyd on the Ezra Klein podcast on my walk home. They talk about a huge range of interesting stuff, but I was especially struck by their discussion of media's role as a filter, and danah's connecting that to how algorithms act as a filter. I think she references "media as the original algorithm" (that's quoted off my memory of the podcast), and talks about how the concerns and critiques we have of algorithms (like Facebook's newsfeed) can also be made of institutions that make decisions editorially.

  2. For me this connected to the recent Propublica article on Facebooks guidelines for human censors. Something about this terrible system they had set-up seems really important and suggestive to me. I still can't really articulate it, but I think it has something to do with the interplay involved between algorithms and people. The result of their policy, shown most clearly in the "protect white men but not black children" slide, is exactly the type of result I would expect from a situation where people had carelessly set-up the rules for an algorithm without really thinking them through. But I'm used to thinking of those terrible systems as the result of the fact that people had to formulate their rules so that a computer could understand them. It's still the people's fault for being careless, but I generally think of a large part of that fault as being a failure to carefully think about how you translate from human to computer.

    At least in that case you have this hope that by talking more about what algorithms are and aren't good at, you can educate people out of their carelessness. But in this case, the algorithm wasn't created to be understood by a computer, it was set-up to tell people how to act. It seems like that means the problem is even larger, that it's a systematically careless and context-less way of thinking that is being used not just because it's the easiest way to have a computer do the job, but because it's the way the people involved think. It gets into that territory of the language you speak defining the things you have conceive of, where maybe Facebook people have been fitting their thought into computer-friendly patterns so long that it has limited their ability to think outside those patterns. (It could go th other direction to, that people who already think that way are going to be attracted to work where thinking that way is relatively effective, so you get a concentration of people who think that way without people who think differently being there to call it out. Or it could be a combination of both things. Part of why I think and worry about that stuff is I worry those people could be me.)

In both cases I think I was especially interested because through work I have been thinking a lot about how and why computer algorithms can fail us. Here were cases that had the same shape as the ones I've been thinking about, but they weren't specifically about computer algorithms, they were about a much broader category of human behavior. And it seemed important for me to think about that some more.

The Project Idea

So those two things are sloshing around in my head, and then I started thinking about possible ways to explore them more. The one I'm interested in thinking through in this post follows pretty closely from danah's comment on the media as a filter algorithm.

One version of the idea is:

  • Train a model on headlines from the New York Times front page. That will get you a model that can tell you how likely it is a given headline would be on the front page. Then use an interpretability tool (which we've been work with at FFL) to inspect that model and see what kinds of things get on the front page.

What's really intersting to me with this idea is that you're making the computer model as a means of better understanding the criteria behind these human editorial decisions. This was pretty directly inspired by a comment Ezra made about it actually being difficult to determine the criteria for which stories get coverage and which do not. That certainly you have some idea about how different aspects of a story (who something was said by, where it occurred) contributed to it's coverage, but it would be very difficult to really pin down. I think this is just because of the scale of all those editorial decisions is difficult for us to hold in our head at one time, as context for the current story. Computers are good at that kind of thing, so we can use them to build that context through training and then interrogate it.

Now, there are lots of holes, or at least things that need a lot more elaboration, in my description on the project. My plan is to write another post walking through how I would think about whether this is really a feasible idea or not, what sort of data you would need to do it, and what sort of methods might be good. But in the interest of trying to actually regularly post things to this blog and thinking in public I'm putting these preliminary thoughts up now. Let me know if you have thoughts, on Twitter or my email is on the info page.