This year’s AI for Social Good seminar brought together 22 experts in artificial intelligence and international development to explore the power of AI and Machine Learning as tools for impact. The goal was to establish common ground, form collaborations, and workshop through concrete problems in a hackathon.
The first two days of the seminar focused on ‘cracking each other’s code’, to understand what each field is doing, scope each other’s jargon and learn which machine learning technique can be applied to each of the case studies brought to the table. Presentations of various success stories highlighted
The success stories also demonstrated the progress made in the field of AutoML. The increasing availability of computing power, coupled with a democratization effort, has made machine learning tools and methods publicly available and accessible to less technically oriented audiences.
The second half of the seminar was dedicated to hackathons focused on specific case studies brought by NGO’s and organizations working in the humanitarian and development fields. We brought a challenge focused on education data and geospatial interpolation.
Case study: can we use machine learning techniques to predict school dropout in Rwanda?
The challenge we set ourselves was to explore the use of machine learning methods to model development outcomes, in this case dropout, in areas for which no survey data was available.
Laterite teamed up with Ruby Sedgwick and Jose Pablo Foch from Imperial College London, and Raghu Rajan from the University of Freiburg, to work on a proof-of-concept for this geospatial interpolation idea.
The data we used in the hackathon focused on primary and secondary school dropout rates for various villages in Rwanda. We combined disparate sources of data that captured dropout at an individual level, socio-economic characteristics of the child’s household, and data from our geospatial backbone. The Laterite geospatial backbone contains information about Rwanda on a grid of up to 100×100 meter cells. All information is collated from publicly available data, and includes features such as soil type, night light from satellite imagery, distances to major roads and village centers.
Could we use machine learning methods to estimate primary and secondary school dropout in villages for which no data is available?
The first results, after just 2.5 days of hackathon do indeed look very promising. We found that gradient boosting decision trees predict reasonable accuracy. Moreover, it was striking how the most important model features were consistent with findings from the existing literature on school dropout in Rwanda. For instance, wealth indicators at household level turned out to be an important predictor. Similarly, night lights observed by satellites in a certain villages or location on the geospatial grid (which could be construed as a proxy for wealth) yields a lot of predictive power.
Given these initial results and the collaborative atmosphere during the hackathon, the next steps are very promising. We hope to continue to work with Ruby, Jose Pablo and Raghu to assess how these approaches can be used to predict other outcomes, and to strengthen future collaborations and partnerships between Laterite and the AI for social good participants.
—
Our colleagues Rik Linssen and José Rubio Valverde attended the AI for Social Good seminar held at Schloss Dagstuhl in Germany (Feb 27-Mar 4 2022).