Graph Learning and Spectral Clustering in High Dimensional Data

Abstract 

In this talk we will discuss about efficient and accurate algorithms for the estimation of graphs from high dimensional data, and for subsequent community detection tasks. Initially, we examine a performant precision matrix estimation method, based on the sparse quadratic approximation of the l1 regularized Gaussian maximum likelihood. Its capabilities are extended to the retrieval of graphs of only non-negatively correlated variables, or equivalently, the problem of sparse adjacency estimation.  Last, we discuss nonlinear reformulations of spectral graph clustering which promote sharp indicator vectors that correspond to optimal graph cuts and improved clustering assignments. The advantages of all introduced algorithms are showcased in a series of comparative tests with the current state-of-the-art on artificial datasets, and their real-world applicability is demonstrated with numerical experiments on medical, image, and financial data.
 

Speaker Bio

Dimosthenis Pasadakis is a postdoctoral fellow at Università della Svizzera Italiana (USI) in Lugano, Switzerland, where he is currently leading the project "Directed acyclic graph partitioning for scheduling tasks". He holds a PhD from USI in Computational Science, and a BSc in Physics from the Aristotle University of Thessaloniki. The focus of his research is centered around algorithms for graph learning and combinatorial optimization for graph clustering and anomaly detection. In the private sector, he is the chief operating officer of Panua Technologies, a company developing customized high-end software solutions for large-scale applications. His work has received several distinctions, including a best paper award at Springer MedGU'22 and outstanding paper awards at IEEE HPEC in 2023 and 2024.