Probabilistic Machine Learning

Assoz. Prof. Dr. Nikolaus Umlauf, Principal Investigator Mag. Johannes Seiler, PhD, Co-Investigator Univ.-Prof. Dr. Stefan Lang, Co-Investigator in collaboration with Dr. Kenneth Harttgen This project aims to better explain the problems of childhood malnutrition in low- and middle-income countries through probabilistic machine learning and to contribute to the monitoring of the Sustainable Development Goals (SGD), proposed at the United Nations Conference on Sustainable Development in Rio de Janeiro in 2012. Recent literature emphasizes the high heterogeneity at both the national and sub-national level, and fo- cuses on identifying drivers of malnutrition with flexible regression models. While incorporating complex modelling approaches, the applied methods are not sufficient to account for all important interactions, i.e., certain factors remain undetected, that could make a significant contribution to the overall situation. We aim to significantly improve monitoring through: (a) an improved database and (b) development of new algorithms for non- standard interactions that will be embedded in the framework of fully probabilis- tic distributional regression models. The novel algorithms will be based on ideas from machine learning such as decision trees (and random decision forests) and stochastic gradient descent type algorithms, suitable for very large data sets. The presented methods can be used for a variety of applications. The modeling approach focuses on the decomposition into main effects and (possibly) complex but interpretable interactions. The new algorithms are extremely memory efficient (including variable selection) and can be applied to virtually any number of observations on a conventional computer. With the methods developed so far, it is not possible to compute such large probabilistic models. Therefore the methods are also very useful for other applications, e.g., in the field of meteorology, real estate modelling, ecology, medicine, etc. 1

Probabilistic Machine Learning Worldwide, millions of children are affected by malnutrition, especially in countries with limited access to healthcare, clean water, or sufficient food. A common consequence is anemia, or low levels of hemoglobin in the blood. To understand the scale of this problem and take effective action, reliable and meaningful data are essential. In our research project funded by the Austrian Science Fund (FWF), we developed an entirely new statistical algorithm that allows complex health data to be analyzed in a fundamentally new way. Our method can simultaneously process millions of data points, from health surveys, satellite imagery, climate and environmental indicators, and socioeconomic variables, and automatically identifies which factors are most relevant. What makes our approach unique is that our models do not just predict average values (such as the average risk of anemia), but the entire probability distribution. This means that for every region, we can estimate not only how likely anemia is, but also how uncertain the prediction is, or, for example, how high the risk is for particularly severe cases. This allows us to identify areas where the health situation is especially unstable or uncertain. One of the central outcomes of our project is the creation of high-resolution maps for over 50 countries. These maps show, often down to the village level, how severely children are affected by anemia. One striking finding: Differences within countries are often greater than differences between countries. These insights help design aid efforts that are more targeted and equitable. Our new algorithm also makes it easier to explain complex relationships, such as how poverty, climate extremes, and inadequate healthcare systems reinforce each other. Our methods are openly available, have been published in leading international journals, and are freely accessible to other research groups. Our project demonstrates that modern statistical methods can uncover hidden patterns in large datasets, creating a solid foundation for better, evidence-based decision-making.