【Team】 Data-Driven Experimental Design Team
【Date】2026/January/6(Tuesday) 14:30-15:30(JST)
【Speaker】Talk by MdAshraful ISLAM, Nagoya University
Title: Dependencies in Learning Graphical Models
Abstract:
We introduce two new methodologies that advance mutual information estimation for mixed-type variables and the identification of non-linear causal relationships,addressing key challenges in data analysis and causal inference. First, we develop a mixed-type extension of the Chow–Liu framework that constructs a dependency forest and estimates mutual information via copula-based joint density estimation and WBIC-based free-energy computation, improving upon conventional likelihood-based approaches and enabling applications such as linking gene expression with SNP data. Second, we propose a non-linear causal discovery method that integrates generalized additive models (GAMs) with the Hilbert–Schmidt independence criterion (HSIC), allowing flexible estimation of additive-noise structures without restrictive parametric assumptions. We provide theoretical analysis establishing consistency for causal order identification, and experiments demonstrate improved causal structure recovery compared to existing methods.
