Room “Sala Seminari” - Abacus Building (U14)
Causal Discovery via Transformed Low-Rank Quantile Surfaces
Speaker
Thong Pham
Associate Professor, SANKEN, The University of Osaka
Abstract
How flexible can a causal model be while still distinguishing cause from effect? Many approaches to causal discovery restrict conditional distributions to changes in location and scale, possibly after a nonlinear transformation. This talk introduces Low-Rank Quantile Surfaces (LRQS), a framework that allows richer changes in conditional distributional shape. Its central assumption is that, in the causal direction, the conditional quantile surface becomes low rank under an unknown monotone transformation. This includes location-scale and post-nonlinear heteroscedastic noise models as special cases, while extending beyond their restrictions. I will discuss the identifiability results underlying this approach and explain how they motivate a nonparametric causal score based on jointly fitting the low-rank structure and the transformation. Synthetic and benchmark experiments illustrate the benefits of this additional flexibility when causal mechanisms involve complex distributional variation or nonlinear observation distortions. The talk is based on joint work with Ryo Kamimura (NeurIPS 2026).
Short bio
Thong Pham’s research focuses on the statistical foundations of causal discovery and the development of practical methods for learning causal relationships from observational data. His recent work addresses hidden variables and complex nonlinear mechanisms. He received his Ph.D. in Statistics from Osaka University in 2017 and is currently an Associate Professor at SANKEN, The University of Osaka. He also holds appointments as a Specially Appointed Associate Professor at Shiga University and a Visiting Scientist at RIKEN AIP.
Contact person for the seminar: fabio.stella@unimib.it