We study theoretical and applied techniques for building interpretable machine learning models. We also research how to extract explanations from black-box pre-trained models (e.g., deep networks) via concepts, counterfactual explanations and interventions.

Constructive preference elicitation is a technique for learning user preference models over combinatorial domains.

Temporal networks generation is a complex task. In particular, we are working on generating temporal graphs that are statistically similar to a given temporal network.

Egocentric Temporal Motifs Miner is a novel mining technique based on the node egocentric perspective. Egocentric Temporal Motifs are stored in a unique binary string; thus, we can verify isomorphism in linear time.

Learning Modulo Theories is a machine learning framework to perform discriminative structured-output prediction over hybrid constrained domains.

Weighted Model Integration is a method for solving inference problems over arbitrary probabilistic models over constrained domains.