Using Machine Learning To Improve Targeting Of Humanitarian Aid

Let's take a look at how Machine learning and phone data can improve targeting humanitarian aid

As cell phones have grown increasingly prevalent worldwide, with a projected global penetration level of 73 percent in 2020, research on wealth forecasting from digital trail data has concentrated on mobile phone metadata (GSMA, 2017). Machine learning algorithms based on call detail records (CDR) have recently been proved to yield meaningful estimations of prosperity and well-being at a fine geographical resolution. Machine Learning and Artificial Intelligence can be used to target poor populations effectively for humanitarian aid using digital indicators.

The challenge of assessing who is qualified for humanitarian help and who is not is a key cause of problems in anti-poverty programme management. Typically, programmes target people based on administrative records like tax records or survey-based asset or consumption measurements. However, the quality of this data is rapidly deteriorating, and trustworthy data for targeting does not exist in many poor nations and will be extremely expensive to gather. However, during the last several years, a number of studies have demonstrated that non-traditional "digital trail" data—behavioral data—can be useful.  Wealth is predicted by indications captured in regular encounters with technology. In environments that are still evolving Machine learning and development groups are hopeful that this data will give a speedy and low-cost alternative to regular data.

Means tests are used in most focused anti-poverty initiatives in the developed world, limiting programme benefits to people who fall below a specified income or consumption criterion. Means tests, on the other hand, are sometimes impracticable in the developing world, particularly in places where the majority of work is now in the informal sector and records of revenue and costs are few. As a result, most poverty-targeting programmes in the emerging world rely on proxy wealth metrics.

Study of using Phone data for targeting in Togo

Machine-learning techniques were applied to high-resolution satellite photos to provide accessible micro-estimates of a relative richness of each 2.4 km by 2.4 km area in Togo. The relative value of all the homes in each tiny grid cell is estimated using these calculations.

The relative value of all the homes in each tiny grid cell is estimated using these calculations.

Machine-learning methods were used to estimate each mobile phone subscriber's average daily use using mobile phone metadata given by Togo's two mobile phone carriers.

This data was utilised to develop supervised machine-learning algorithms that can predict wealth and consumption based on phone usage.

While simulated targeting of a theoretical national anti-poverty programme, a high-dimensional variable of mobile phone features was used instead of a low-dimensional variable of assets to estimate wealth, and machine-learning algorithms were used to maximise out-of-sample forecasting accuracy instead of the conventional linear regression that maximises in-sample goodness of fit.

The findings were given, which compared the efficiency of these various targeting techniques in two separate scenarios. First, we look at the actual policy scenario that the Togo government faced in September 2020, which required providing cash to 60,000 people in the country's lowest 100 cantons. The data acquired in a big phone poll we built for this purpose and performed in September 2020 is used to evaluate the first scenario. Because consumption could not be acquired realistically in the phone survey, the 'ground truth' measure of poverty for this first scenario is a PMT. The PMT is built on a progressive regression approach that captures around 48% of consumption variance. Second, we simulate and assess a more general & hypothetical policy scenario wherein the government is concerned in focusing on the poorest people across the country; this scenario is assessed using government-collected national household survey data from 2018 and 2019. In the second simulation, the ground truth definition of poverty is consumption. These data are presented in the 'Data sources' portion of the Methods section, and the assessment is detailed in the 'Targeting evaluations' section of the Methods section.

When the demographic under consideration is more homogenous, and when other characteristics (such as residential location) that are employed in more traditional ways to targeting are less variable (Techniques, 'Targeting methods and counterfactuals'), phone-based targeting is most effective. For example, when the simulation of a hypothetical national programme is limited to rural homes, the benefits of phone-based targeting grow.

The fact that the machine-learning model was built on representative poll data acquired just before the program's rollout, was a key component in its success. Because a person's poverty status may vary over time, and the best phone-based indicators of wealth might change as well, a model trained for one year or period may not perform well when used in another year or season.