This open-source tool developed in collaboration with the IDB applies machine learning to review the quality of data used to determine eligibility for social programs in Colombia and the Dominican Republic.

One of the keys to implementing or evaluating a social program is having good data. And, in this sense, well-designed information systems for social programs are key to their effectiveness. These programs depend on massive information-gathering exercises related to households, in order to determine eligibility to receive benefits connected to health, education or conditional transfers, for example. However, the surveys typically employed to gather this household data do not always allow for manual verification for each of the variables that are asked. This certainly leaves rooms for errors or atypical data in the information obtained. For this reason, improving the quality of this data can in fact strengthen the quality of social program delivery, allowing us to provide services with greater precision and efficiency.

As in so many other fields, the digital revolution is making it possible to apply technology including artificial intelligence to obtain better results in social policies. A new open-source application that applies machine learning techniques to improve and accelerate the revision of this data is an example of this. The Atypical Data Classifier, formerly known as the Identification System for Potential Beneficiaries of Social Programs in Colombia (SISBEN ML) is a system that was designed with the purpose of automating a quality control process, taking into account all available information of household surveys in an objective way to select the cases that merit verification. This tool automatically classifies atypical cases of information to improve data quality and efficiency in the review process of potential beneficiaries of social programs.

The tool was intentionally developed in open source and is available through Code for Development, an Inter-American Development Bank initiative to promote using open-source technology in Latin America and the Caribbean for the public good. This is a result that we hope to continue replicating at the IDB as we collaborate in the development of other IT solutions together with countries.

Classifying and visualizing social program data

Traditional processes for reviewing social program data use logical validation meshes or manual validation systems. For example, a logical validation mesh verifies that someone born in 1975 cannot be considered a child in the database. The manual review depends on a person reviewing and cross-checking each of these points among possibly thousands of household surveys one by one.

The Atypical Data Classifier uses machine learning to automatically review all the information available from the survey while also potentially flagging atypical cases that may be less obvious, such as the use of an unusual construction material in the area. The current version applies unsupervised machine learning to generate the classifications, meaning the system continually learns by itself what is atypical and what is not. This learning is contextualized by geographic area. The algorithm knows to classify cases based on the local conditions where the surveyed family resides.

The Atypical Data Classifier includes two components. One is the classifier itself, the algorithms used to classify the households by reviewing the survey data and monitoring to detect atypical cases. The second component is the visualizer, a Web interface that allows you to see which cases that the classifier has identified as atypical while highlighting the exact variables within each form that it considers to be inconsistent. This streamlines the process of reviewing the household indicators for possible errors.

Numerous advantages to automating the data quality review process

Automating this process has a series of advantages for the institution that manages social programs:

1. First, it allows you to reduce costs by minimizing the personnel required to review the data collected.
2. It also allows to increase the quality of the final database. In the manual case, random sampling would be carried out in order to deliver the results on time and keep costs down. In this case, it is the algorithm that selects the sample after making a first revision to all the files.
3. Finally, the tool allows to correct any logistic problem or the survey tool during the execution of the operation, since the algorithm will yield in real time the results of the analysis of the data deposited in the central database.

The tool was developed in coordination with the National Planning Department of Colombia, and can be adapted for use in other countries. Any organization can calibrate the components with the weights that they value in their calculations for determining eligibility for social programs.

In the process of creating the tool we learned a lot about the types of validation that are necessary for a tool to be reusable by third parties. We are currently working on a more modular version that has fewer dependencies with technologies from a particular provider. This would allow any institution to use the system, regardless of its infrastructure, operating system and other tools that it might be using.

In the near future we hope to take an additional step through the use of information from social databases and tools that apply machine learning to further enhance information and improve the efficiency of the use of public resources in social programs.

Interested in using this tool? Get the code for the Atypical Data Classifier on our GitHub.

 

By Luis Tejerina, Lead Specialist in the IDB Social Protection and Health Division and Carlos Tejada, Information Systems development consultant