Computational Social Sciences
I am interested in developing statistical and machine learning tools for analyzing large datasets and addressing computational problems that arise in applied sciences.
Machine learning methodologies can be adapted to a wide and heterogeneous spectrum of applications in social and health sciences. The sort of problems I have dealt with regard:
policy prediction problems;
new tools to enhance interpretability of machine learning algorithms.
1. Machine learning techniques can be exploited to answer social science research questions that are directly related to prediction. Massimo Riccaboni (IMT School for Advanced Studies), Armando Rungi (IMT School for Advanced Studies) and I worked on improving over current methodologies in predicting the risk of failure of Italian firms and to provide a novel definition of zombie firms (i.e., non-viable firms that manage to stay on the market despite low productivity and poor financial shape). In this work, we have proposed to use a new Bayesian machine learning methodology that improves over alternative supervised learning techniques in precisely predicting the failure risk of Italian manufacturing enterprises between 2008 and 2018 (for a comparison see our recent review of the literature on supervised learning for firm dynamics).
2. Your machine learning model performs accurate predictions. But can you safely assume that you are not missing any important predictor in your model? How the inclusion of a new variable with high explanatory power would change the predictions of your model, and, in turn, its performance? Together with Kenneth De Beckker (KU Leuven, UHasselt), Kristof De Witte (KU Leuven & Maastricht University) and Joana Maldonado (KU Leuven) we developed a new sensitivity analysis for machine learning methodologies to answer such questions. We applied this new technique to a prediction problem related to correctly forecasting students financial literacy scores in a region of Belgium where such scores were not observed.