How AI is being used to map poverty

According to the World Bank, 736 million people live in extreme poverty worldwide, and half of them are in just five countries. These are (in descending order) - India, Nigeria, Democratic Republic of Congo, Ethiopia, and Bangladesh.

Among those leading the fight against poverty are the World Bank, one of the world’s largest sources of funding for poverty reduction projects, and the United Nations (UN) - whose number one priority among its 17 Sustainable Development Goals is to ‘End Poverty in all Forms Everywhere’. The UN also holds the International Day for the Eradication of Poverty on 17th October.

Both the World Bank and the UN rely heavily on data and research to measure progress towards this goal, and the Qatar Computing Research Institute (QCRI), which is part of Hamad Bin Khalifa University, are collaborating with several such organisations to tackle global problems using Artificial Intelligence (AI).

QCRI has a long history of engaging with UN agencies and NGOs. For example, it has developed the Artificial Intelligence for Digital Response (AIDR) platform, which received the Open Source Software World Challenge Grand Prize in 2015.

AIDR analyses data during disasters, such as the recent Hurricane Dorian, and the QCRI work directly with relief organisations around the world to develop the technologies they need to make sense of ‘big data’ during a disaster to better assess the situation and allocate resources. The QCRI are now also using machine learning to produce poverty maps using anonymous advertising data from Facebook.

How a person’s mobile device can determine wealth

Facebook provides advertisers with the capability to selectively target their message to users matching certain criteria. The supported targeting criteria includes age and gender, but also attributes such as the countries a user has lived in or the device used to access Facebook. To facilitate budget planning, Facebook provides so-called audience estimates, i.e. estimates of how many users match the provided targeting criteria.

For example, according to Facebook, there are 2.4 million users aged 18 and above living in Qatar. Of these, 280,000 lived in Nepal and 12,000 lived in the US. Of the users who used to live in Nepal, only four percent use an iOS device as their main method of accessing Facebook, whereas the percentage is 38% for users who lived in the US. This kind of analysis can even be done at the sub-city level showing, for example, that users from Western countries are more likely to live in more expensive areas such as West Bay or the Pearl, whereas users from Nepal are more likely to live in the Industrial Area.

As, generally speaking, iPhones are more expensive than other types of mobile devices, this observation indicates that, in Qatar, Facebook users from the US have a higher disposable income than users from Nepal, and that this is also reflected in a variation in disposable income across different parts of town.

Of course, many people who could afford an iPhone choose not to buy one but, looking across large enough groups of people, the QCRI find that the device type used to access Facebook can be a powerful signal to map relative levels of poverty.

For readers living in Qatar, the example above will not be surprising. But the power of the general methodology – looking at the device as well as internet connectivity type – comes from the fact that this also works internationally.

Why poverty maps are necessary

QCRI has been using this data for a year to map poverty, in particular in the Philippines and India. The QCRI do this because many countries, particularly developing countries, do not have up-to-date data, which makes it difficult to plan or evaluate effective poverty reduction measures.

The QCRI combine the audience estimates from Facebook with other data sources, such as satellite data and population density estimates, to produce poverty maps. This works by first describing locations in terms of a set of ’features‘, that is a list of numbers summarising both the Facebook audience found in the location as well as how the location appears on satellite imagery.

The QCRI then use available poverty data for locations where it exists to learn a function that computes a poverty estimate for a location from the set of features. Once we have learned this function and evaluated its accuracy, we can apply it to make poverty predictions for locations where no ground truth data exists.

The QCRI started this line of work with the Philippines in collaboration with UNICEF Innovation and a Philippines-based startup called Thinking Machines, initially only using Facebook data to make predictions. Following good results, it was extended to India and data from satellite imagery has been incorporated.

To illustrate how satellite data could be used to map poverty, previous researchers observed that areas with a lot of light at night, as seen from space, have high economic activities. This has been used to build models to predict, say, the GDP for locations such as North Korea, or other locations where trustworthy statistics might not be available. Researchers have since extended such methods to work with daytime satellite data, picking up signals such as whether roofs are thatched or tiled, or whether roads are paved or not.

The QCRI has found that satellite data alone works better in rural areas and Facebook data alone works better in urban areas, but the combined model is even better in both.

Sharing knowledge to make an impact

To go beyond models and to have a real world impact, it is essential to develop partnerships. At this year's AI for Good Global Summit held by the International Telecommunication Union in Geneva, the QCRI organised a half-day workshop on poverty mapping in collaboration with the World Bank and Dalberg Data Insights.

One outcome of this meeting was the creation of an online group of poverty mapping experts. Through these outreach and partnership efforts, the QCRI hope that its work can lead to evidence-based decisions on things such as where to target public investment to improve service delivery and infrastructure.