Hi! 👋
My name is Riccardo Massidda, and I am a post-doc researcher at the University of Pisa, Italy.
My research lies at the intersection of machine learning and causality and focuses on causal abstraction methods.
I completed my Ph.D. as an ELLIS student, under the supervision of Davide Bacciu (University of Pisa) and Sara Magliacane (University of Amsterdam).
Reach me at name.surname [at] di.unipi.it
📜Publications
Learning Causal Abstractions of Linear Structural Causal Models
Uncertainty in Artificial Intelligence (UAI 2024)
This paper provides theoretical foundations for learning causal abstractions between linear models and introduces Abs-LiNGAM, a method that leverages abstract causal information to execution time of causal discovery of linear non-Gaussian models.
Constraint-Free Structure Learning with Smooth Acyclic Orientations
International Conference on Learning Representations (ICLR 2024)
In this paper we introduce COSMO, a constraint-free continuous optimization scheme for acyclic structure learning. In addition to being asymptotically faster, COSMO compares favorably with constrained approaches.
Causal Abstraction with Soft Interventions
Oral at Conference on Causal Learning and Reasoning (CLeaR2023)
Causal abstraction provides a theory describing how several causal models can represent the same system at different levels of detail. In this work, we extend existing causal abstraction proposals to soft parametric interventions.
Knowledge-Drive Interpretation of Convolutional Neural Networks
Oral at European Conference on Machine Learning (ECML 2022)
Definition of a framework to align ontological concepts and neural directions. Design of ablation experiments to validate the causal role of neurons within semantically coherent clusters.
rmassidda@ DaDoEval: Document Dating Using Sentence Embeddings at EVALITA 2020
7th Evaluation of NLP and Speech Tools for Italian (EVALITA 2020)
Definition of a hierarchical architecture based on sentence-embeddings (USE, LaBSE, SBERT) to solve a document dating task at different granularity levels.
Technology and Species independent Simulation of Sequencing data and Genomic Variants
IEEE 19th International Conference on Bioinformatics and Bioengineering (BIBE 2019)
Development of a data-driven bioinformatic model for the simulation of genomic variants and next-generation sequencing (NGS) reading errors.