Demystifying Machine Learning for Global Development – Stanford Social Innovation Review
Machine learning is an increasingly prevalent buzzword in the media. Its applications in science and the private sector are frequently discussed—but what about global development? Can it also help advance fields like health, agriculture, and financial inclusion?
Absolutely. That’s because it can help us uncover previously invisible patterns in data, to identify the most effective solutions and target them in the right way.
Machine learning (ML) has been around for decades, but now is our chance to apply it to development challenges in new ways, for three reasons. First, ML has advanced in recent years; better algorithms and open-source software have made ML tools widely available and accessible. Second, the infrastructure to manage, share, and analyze data (including high-speed computational power) has become affordable at scale. And third, there’s been an explosion in the amount of available development data, thanks to both deliberate data collection efforts like surveys, program monitoring, and evaluation studies, and new data sources from satellites and mobile phones.
At Surgo Foundation, we’ve been applying ML to several development efforts, including increasing the number of women who deliver babies in health facilities, and improving the performance of front-line health workers to reduce maternal and child mortality in India. We’ve seen ML spur innovation and believe that we and other global development organizations could be getting much more out of our data and using it to greater effect.
Nevertheless, ML can be a daunting concept, hard to grasp, and technically overwhelming. The good news is that there are several types of ML approaches, and they’re not all as complex as those that get the most media attention.
The best way to approach ML is to identify how certain types of development questions might benefit from the application of ML, as opposed to discussing ML in the abstract. Below are four such questions, and a look at the readily accessible and applicable ML tools that can help answer them.
1. How can written information better support our decision-making?
The development sector generates huge amounts of text in the form of medical records, published research, government records, and more. The vast volume of information makes it untenable for humans to analyze efficiently, undermining the valuable insights and trends that can be generated from it.
Natural language processing is a kind of ML algorithm that draws insights from large amounts of text. It can streamline and improve clinical processes. For example, multiple health care providers often see the same patient but generate separate clinical records. ML can extract and analyze keywords or concepts from these separate records so that providers can develop a holistic and accurate diagnosis. Natural language processing can also translate “free” text, such as health providers’ notes, into structured data for analysis. Yet another application is conversational agents, or chatbots, which interact with people by interpreting the text they send through their mobile phones. Chatbots have huge potential to aid medical diagnosis and provide information to people who live far from health services.
Machine learning can help us uncover previously invisible patterns in data, to identify the most effective solutions and target them in the right way
Organizations can also use natural language processing to analyze patterns, such as trends and gaps in funding and research by donors and governments. For example, the World Bank’s Project DeCODE uses it to analyze written documents, and glean information that can help predict changes in development projects and guide decisions about new investments.
2. How can we provide the right person with the right intervention?
People are not the same. To achieve impact and shift behaviors, development organizations need to move away from a one-size-fits-all approach and toward interventions targeted to different sub-groups of the population.
Cluster analysis is a powerful ML technique used widely in the private sector to segment customers for targeted marketing. The basic principle is to group people in such a way that group members are as similar as possible to each other and as different as possible from those in the other groups. Development programs typically base their clustering efforts on demographic factors such as age, gender, education, or urban-rural residence. But the beliefs, motivations, biases, and norms that underlie people’s decisions, as well as potential structural barriers (such as the availability of a service), are equally important to targeting efforts and driving change.
An interesting case study of clustering in development is a global program to encourage voluntary medical male circumcision to help prevent the spread of HIV. To help governments in Zambia and Zimbabwe increase demand for this service, our foundation conducted a large-scale survey to quantify the beliefs, biases, and community norms that influenced men’s attitudes and behavior toward circumcision. The data on these variables went into the clustering algorithm. In Zambia, this ML approach gave us seven groups of men with different beliefs about circumcision, which we then used to develop more-effective and targeted interventions.
3. How can we anticipate events, behaviors, and market dynamics?
Being able to predict what will happen, where, and when can help development organizations focus their limited resources on the right interventions, people, place, and time.
Predictive machine learning helps achieve exactly this. There are two main types: classification and regression. Classification predicts which target class a person belongs to, such as whether a pregnant woman is likely to have a low-risk or high-risk pregnancy. Regression predicts a specific value, such as the income of an individual, given a set of demographic and geographic indicators.
Classification methods can, for example, identify poor households in a community using minimal amounts of national survey data, whereas using traditional analytical methods would be expensive and time-consuming. Organizations are also using classification to develop mobile phone-based health diagnostic tools. In the example of voluntary medical male circumcision above, ML made it possible for community health workers to predict which segment a man belonged to using his responses to a few simple interview questions.
Regression models have a variety of applications too. They can predict levels of wealth based on satellite data identifying homes and property, for example, or the amount of corruption based on administrative indicators such as government data on financial transactions. In the health sector, regression models can simulate and forecast the market share of different contraceptives—injectables, condoms, oral contraceptive pills, or intrauterine contraceptive devices—based on customer preferences so that governments can better plan how much of each product type they should purchase and distribute.
4. How can we understand cause and effect?
In development, we are obsessed with causation. Our default approach to this has been using randomized control trials, but they can be costly, take years, and are often only able to test the effect of one intervention at a time.
Causal machine learning lets us identify the network of factors that influence a development outcome. This moves us beyond making predictions, to helping us understand the underlying causal relationship between variables.
For example, to truly understand infant mortality in a given setting we need to know more than which factors might correlate with the death of a baby. Instead, the question ML can help answer is precisely which characteristics, behaviors, and health parameters of mothers, and which behaviors by health care providers, lead directly or indirectly to infant deaths? ML algorithms can help us map how all these factors interact with each other, and once we define this underlying, explanatory structure, we can do “what if?” experiments with the data. For example, what if we had front-line workers contact pregnant woman five times during their pregnancy instead of three times? How much change in infant mortality would we expect to see?
Development organizations need to move away from a one-size-fits-all approach and toward interventions targeted to different sub-groups of the population.
We are currently testing this approach in India. We are using numerous vertical datasets to try to understand the complete set of factors that explain the outcomes we see in that location. We collected variables both from the mothers, such as how many antenatal checkups she attended, and from the health facilities, such as nurse practices. Putting these variables into a regression model might reveal that a clean umbilical cord or an educated mother is associated with infant survival, but all we would know is that the two are correlated. It wouldn’t necessarily be clear whether these were causal factors, or whether other factors played a role. By contrast, causal models help us understand how all the factors are linked in a network, see which ones impact the outcome, and identify critical touch points in that system to improve survival rates.
Overcoming the Challenges
Some think ML is a silver bullet to solve all our global development problems, but it’s not the answer to every question. Development organizations should recognize it as a valuable tool, one of many complementary approaches to solving a problem. Meanwhile, there are several issues to consider before deciding whether to use ML:
- Quality and quantity of data are still an issue. While global development is now more data-rich, we tend only to collect data we think is important, and we may therefore miss significant predictors or causal factors of behavior. Data quality varies widely, and data sets are often fragmented, as donors or governments each collect data based on what they need. It’s difficult to integrate these data sets if owners are unwilling to share them.
- We need to pull ML talent into the public sector. The sector needs to draw and build ML expertise to make it a common, scalable tool for the sector broadly, just like the statistical tools that many are already use.
- Context is all-important. Development experts must work closely with ML experts to ask the right questions and interpret the patterns. There will also be a need for roles like ethicists and policy experts as ML scales up.
- We need to become good customers of ML approaches. We need a clear understanding of what development problems can most benefit from ML. This doesn’t mean knowing all the stats and algorithms ourselves, but we do need to educate ourselves and each other on the different approaches—on their value and shortcomings. Then we can have productive conversations with ML experts or use the many available web-based services, such as RapidMiner, which make it possible to use ML tools with minimal coding. Finally, we need to build case studies to test the approach and demonstrate what value the application of ML has brought to programs.
It is important to realize that the most impactful techniques for any given global development effort are not necessarily the most advanced ones or the shiny objects most in the news. ML may help us in certain situations; basic statistics or other tools in others. The key is to start with the problem—then find the right tool.