The Humanitarian Space - Multi Armed Bandit Testing

(Mike Nolan) #1

Hey all,

My project has gotten to a point where I have things which are kind of fun to show. So I work for the International Rescue Committee, they are this big organization, kind of like Red Cross, that help people who have been affected or displaced by conflict or disaster. So anyone that’s internally displaced within their own country, or in the case with my work specifically, a refugee, we work to help rebuild their lives.

Anyways, so I’m working in Jordan helping Syrian refugees and economically vulnerable Jordanians find employment opportunities. We have this suite of “interventions” which are like aid packages. Some are cash hand outs, some are information sessions, some are focused on confidence building (we call this our psychological intervention).

We don’t really have any evidence or data if one works better than the others. We don’t know if one works better for certain groups or worse for other certain groups. We essentially don’t know anything at all.

So what do we do? Well, essentially we want to measure the effectiveness of these treatment packages, and shift our program to hand out the things that are most effective. So, you’ll see a lot of websites employ a strategy called A/B testing, where a certain percentage of users get version A of a website, and a certain percentage of users get version B of a website. After some time we check back and look at the data and see if version A is doing better or if version B is doing better.

What we’re doing here is similar but slightly more advanced. We’re employing this strategy called Multi-Armed Bandit Testing. Read the link for a bit more information but the gist is that instead of having to wait some time and look back, we employ machine learning to evaluate the data, and shift the probability of who gets what treatment as it becomes more and more confident as to which treatment works best for which type of person.

We even have this cool little dashboard that we’re using to monitor how the treatments are shifting over time. All of our code is open sourced on github. Given the sensitive nature of refugee data, we’re working on a solution to making multi-armed bandit testing a bit more plug and play in this sort of situation but stay turned for more information onthat.

Does this sort of stuff interest you? If so, you should definitely check out LibreCorps. They work with a similar organization to mine called UNICEF doing great work as well. Always feel free to DM me on IRC (nolski[m]) if you want to chat more about this kind of work.

(Justin W. Flory) #2

Thanks for sharing @nolski, this is on my reading list for this week. :slightly_smiling_face:

More info about LibreCorps, a program run by Prof. Stephen Jacobs at the MAGIC Center is on the FOSS@MAGIC website too if anyone is curious.