Four tips on using machine learning for social good

Children in a village school in Bhilwara District, Rajasthan, India. ©IDinsight/Elizabeth Bennett

The piece summarized below was originally published at Stanford Social Innovation Review, written by Ben Brockman, Andrew Fraker, Jeffery McManus, and Neil Buddy Shah.

IDinsight recently published an article in Stanford Social Innovation Review detailing how non-profits, governments, and others can use machine learning to increase their social impact. The article outlines enabling conditions that can help service providers get the most out of machine learning.

What is machine learning? “Machine learning uses data… and statistical algorithms to predict something unknown. In the private sector, for example, ride sharing apps use traffic data to predict customer wait times.

The featured case study with Educate Girls explains how we used machine learning to identify which geographic areas they should target to maximize the number of out-of-school girls they would find. The result? Educate Girls’ program is expected to enroll more than 1.5x the number of expected out of school girls at similar operational costs.

The article points to four requirements for maximizing machine learning for social impact:

1. Good predictors: This is the data that can help predict the outcomes we are interested in. In the case of Educate Girls, we were trying to understand where out of school girls were concentrated. We used publicly available census and education system data to predict which areas where likely to have the most out-of-school girls.

2. Outcome data: This is the data focused on the outcome of the program we are trying to predict, for example, the number of out-of-school girls in a given village. Educate Girls had already collected enrollment status data from one million households across 8,000 villages. This data, when combined with the predictor data described above, is valuable because it allows us to analyze patterns that link the predictors to the outcome e.g. we could better predict which areas had the most out-of-school girls and which had the fewest.

3. Capacity to act on predictions: A machine learning algorithm can accurately predict which villages or districts are most likely to have large numbers of out-of-school girls, but the implementing partner will need to have the ability to divert resources to these villages or districts. In this case, Educate Girls has the discretion to choose which states, districts, and villages it will expand its program to and can now rely on our algorithm to inform those decisions.

4. Ability to maintain the machine: In order for a machine learning algorithm to continue to make accurate predictions, it has to be updated with fresh data from time-to-time. For example, factors that predict school enrollment in one time and place in India may be different another time and place. As such, it is important that Educate Girls update its prediction algorithm from time to time, particularly as it expands to new geographies. The more thorough and up to date the data, the more accurate the predictions.

The article closes with a discussion of different possible machine learning applications and areas for investment, focusing on agriculture. Read the full article here for a more in-depth discussion of machine learning and its applications to the social sector.

Educate Girls was recently selected as an awardee for The Audacious Project. We look forward to continuing to partner with their team as we push forward on this critical work. Watch this space for updates!