Knowledge for Development: A Machine Learning Approach
The wealth of development data has not been mined or analyzed at scale, limiting our understanding of whether or not investments are achieving competitive results on the ground. However, recent advances in machine learning enable the aggregation of findings from millions of research documents, project evaluations and other key sources, thereby unlocking incredible insights into what works in international development.
One foundational insight enabled by machine learning is a better understanding of project design. By “reading” every project evaluation available, semi-supervised machine learning enables the aggregation of an incredible amount of information. One result of such an effort is a map of the trajectory of interventions across time and space, revealing the details of portfolio composition. When analyzed at scale, portfolio composition is revealed as a sort of Rubik’s cube of interventions, which combine in different patterns across time to form the design of different projects and the scaffolding of the portfolio itself. Such clarity around portfolio composition is a prerequisite for establishing what impact, if any, development projects have.
Taking USAID as an example, Figure 1 below maps portfolio trends over 123 countries and 60 years within USAID’s agriculture portfolio. If Organizational Knowledge Theory, which postulates that organizations learn from past experience, is to be believed, then a fall in the popularity of an intervention would indicate that USAID learned from the failure of that intervention at achieving the desired result and implemented the necessary course correction.
To explore the above visualization of USAID Agriculture Portfolio Trends in detail, visit Tableau Public.
For example, the orange line in Figure 2 below tracks the popularity of the “research and development” intervention across 60 years of agriculture projects implemented by USAID. Research and development was a popular intervention up until the 1990s, but then fell dramatically in popularity around 1995. Could this rise in popularity be related to the success of agricultural research and development during the green revolution? If so, why then did it go out of fashion? An alternative explanation could be that, contrary to Organizational Learning Theory, USAID is not learning from past experience but rather intervention implementation is more related to congressional earmarks or other external factors than to learning from past experience. Further research into the causal mechanisms behind the trends is required to understand the full picture.
As stated by Lord Kevin, “When you can measure what you are speaking about and express it in numbers, you know something about it. (Otherwise) your knowledge is meager and unsatisfactory kin; it may be the beginning of knowledge but you have scarcely advanced to the stage of science.” Machine learning gives us this power to speak in numbers by quantifying a vast amount of qualitative information that hitherto could not have been analyzed by manpower alone. Thus, we are now able to better understand trends and forward our understanding, and ultimately our impact, within international development programing.