Using Machine Learning to Help Do Good
Written by Anisha Chikkareddy
Edited by Justin Gambill
February 9th, 2023
Edited by Justin Gambill
February 9th, 2023
Research
In a world devastated by a raging pandemic, global warming, deepening social divides, and an increasing lack of hope for the future, any help is needed. However, humanitarian aid has a “history of inefficiency … [that] often fails to reach the recipients who need it most” (Harvard International Review). What if there were a way to maximize the efficiency of aid?
With the help of Machine Learning (ML), phone data can be used to target the people most in need of aid. In the study performed by a group of researchers in the School of Information at the University of California, Berkeley, an approach to targeting social assistance that uses data from mobile phone networks was developed. The study was done in Togo, and was based on a current program: Novissi, an emergency social assistance program used in response to COVID-19 (Novissi). Novissi “aimed to provide subsistence cash relief to those most affected” (864). The ML program constructed a poverty score for every individual subscriber, and eligibility for aid was determined by a complex function that was told to target Togolese habitants who needed the most help. This ML approach was tested against three other approaches to targeting: geographic, occupation-based, phone expenditure-based. Geographic targeting gave aid towards the poorest prefectures and cantons; occupation-based targeting gave aid towards informal workers and statistically lower-paying Togo jobs; and phone expenditure-based targeting gave aid to people who spent less on their phone bills (the theory being that a higher total expenditure on calling and texting was indicative of higher wealth). They tested these methods in two different scenarios: the first evaluated the actual policy in Togo, and the second evaluated a more generalized hypothetical scenario.
The study found that in both scenarios the ML approach did much better than the other three, which shows that it is a viable and a notably better alternative to the current, more conventional methods of dispersing humanitarian aid. In fact, ML managed to reduce exclusion errors by 4-21% relative to Togo (Aiken et al., 864). On the other hand, when compared to established programs (globally) that had access to a comprehensive social registry, it unfortunately increased exclusion errors by 9-35% (Aiken et al., 864). While there are some limitations and drawbacks, this new approach can complement traditional methods, and can even better predict what to do in urgent crises where the problem is ever changing and unpredictable. It can help eliminate human error, corruption, and the currently rampant inefficiency. While ML-based humanitarian aid is still far from perfection, the implications of this are widespread and give hope towards a better future.