Tuesday, April 1, 2014

Turkers

In 1770, a wonderful machine was revealed. The Turk (also known as Automaton Chess Player) was a chess-playing machine that beat various famous opponents.

If you've ever heard of "Amazon Mechanical Turks," this machine is where they got their name. Amazon Mechanical Turk is a marketplace where people who are willing to do "Human Intelligence Tasks" choose from available tasks which companies are willing to pay to have done. These tasks are generally things computers are not great at -- classifying a sentence as funny or not for instance -- but that humans can do easily. As you probably guessed by now, the original Turk was not actually a machine, but in fact had a chess master hide inside it. I highly recommend reading the wikipedia link, how elaborate the hoax was made my day.

So how does this relate to decision making? First off, you might use the data directly to make decisions. However, while computers may not be good at determining if a sentence is funny, you can train them on data to predict whether a sentence is funny. And it turns out you can use Turkers to generate that training data! Kartik Hosanagar used Natural Language Processing algorithms in conjunction with Turker-generated data on various corporate Facebook posts to attempt to infer what drives consumer engagement (the abstract is available here). This seems like a really cool technique which can be used to inform a whole lot of decision making in the future.

Wednesday, March 19, 2014

Liver Allocations

Today my department had Sommer Gentry give a presentation on her work on liver transplants, an area I previously knew very little about. The basic idea is that because livers only last up to 8 hours after the donor dies, we can't simply allocate to whoever in the country needs it the most. Instead, historically there were 59 DSA (districts) and 11 regions (collections of districts). So within each DSA, an available liver is allocated to whoever needs it most.

This seems like a pretty good way to handle things until you know that liver availability does not necessarily match up with demand. Sommer's research has to do with coming up with collections of districts which will lead to a more equitable allocation across the US.

The big takeaway to me of her presentation was that transparency is key if you're actually hoping to have policy change because of your optimization results. She explained that some things which you would naturally want to have as an objective (distance organs travel for instance), end up as constraints when you want a model you can explain to anyone. This was particularly important since, as you can see in the video below, some entities are major losers in a more fair world. If the stakeholders can't find a problem with the process, just the outcome, then it will be a lot easier to achieve change!

I really enjoyed her presentation, and encourage you to watch the entire video (This seems to be a more mathematical version of the presentation I saw). The jump is set to her presenting a comparison of her solution vs. the current one in a way that visually shows the difference in fairness, which hits on transparency once again.

Feel free to post questions about the problem below!