I obtained my bachelor’s degree in math at University of Trento. For my master’s degree I moved to Torino where I continued my studies in math focusing on probability and statistics. During my period of study abroad at University of Waterloo I discovered machine learning and I decided to make it my main field of study. I graduated with a master’s thesis on the COVID-19 outbreak with Professor Andrea Pugliese. After graduation, before landing in Oslo, I joined the SALSA team at University of Bolzano.
Research interests and hobbies
My research mainly focusses on machine learning both supervised and unsupervised. In particular I am interested in explainability of the models. I also worked with some algorithms in reinforcement learning. The main topic of my PhD will be boosting algorithms applied to graph-structured data.
Aside from my mathematical research I enjoy climbing, now that I moved to Norway I am also looking forward to start cross country skiing.
Statistical learning method for chemistry applications
The driving goal for this project is to create a supervised framework that helps us to understand non standard structured data.
With this perspective this project has two main objectives:
- Refine existing approaches, such as boosting algorithms or neural networks, to include domain-specific knowledge, to improve their performances.
- Develop measures of uncertainty around their results, to allow the users estimating how much they should trust the results.
The latter, in particular, is a topic of extreme interest in all the fields in which statistical/machine learning is implemented and goes in the direction of explainable AI.
During the studies we will team up with the research group at Hylleraas Centre for Quantum Molecular Sciences (Dr. David Balcells) who granted us access to a large dataset of transitionmetal complexes.This dataset allows us to use non standard-structured data. Indeed the molecules recorded are organized into graphs. This will force us to study a new way to explore the large space of graphs that can be used to reppresent the molecules.
The dataset provides also a set of labels based on quantum properties of the molecules. With this data the goal is to predict this labels firstly basing our choice only on the graph structure. In a second time it could be possible to add a new layer to our model that will allow us to include also the atoms attributes in the decision process.
The project also includes an external participation of the University of Bonn (Germany).