Child Poverty Measurement: A machine learning approach

Child poverty is still a severe issue. One of the UN Sustainable Development Goals (SDG) is to end child poverty by 2030. In order to achieve this goal, it is crucial to target poor children and help them accurately. Nevertheless, child poverty is complex. There is no one-size-fits-all child poverty measurement. Currently, there are three most popular methodologies of child poverty measurement. They are Multidimensional Poverty Index (MPI), Bristol Approach and Multiple Overlapping Deprivations Analysis (MODA). These methods are complicated. To implement these methods need expertise and much time. Besides, for each new dataset, the processes must be implemented again. In this thesis, a machine learning approach to child poverty measurement is developed. The training data is from DHS Guinea 2018. The models of neural network, gradient boosting, random forest, linear SVC and KNN are built and trained. After a comparison of the accuracy and ROC of the models in the test dataset and cross-validation, the models of gradient boosting of all three age groups achieved the best performance. Three best models after feature selection according to three age groups (0-4 years old, 5-14 years old, 15-17 years old) are chosen. By employing a machine learning approach, with the trained machine learning model, it can measure child poverty automatically with new dataset.