Learning and clustering tasks on graphical structures

Decanato - Facoltà di scienze informatiche

Data: 16 Novembre 2022 / 12:15 - 14:00

USI Campus EST, room D5.01, Sector D

Speaker: Dimosthenis Pasadakis, Università della Svizzera italiana

Title: Learning and clustering tasks on graphical structures.

Abstract:  Estimating the graphical structures of high dimensional data and identifying the presence of clusters in them are ubiquitous challenges in every scientific domain that deals with interacting or interconnected variables. In this talk we examine efficient and accurate algorithms that learn and cluster graphs. Initially, we recap the development of a performant precision matrix estimation routine based on the sparse quadratic approximation of the l1-regularized Gaussian maximum likelihood method. Comparative results with various modern open-source packages reveal that this method accelerates the computation of sparse inverse covariance matrices by several orders of magnitude, while attaining equivalent accuracy scores. Additionally, we demonstrate the capabilities of this approach to retrieve graphs of only non-negatively correlated variables, and introduce two algorithms for sparse M-matrix estimation applicable in large-scale real-world scenarios. Finally, we present the utilization of such graphs in clustering tasks, with a nonlinear reformulation of the spectral method in the p-norm that is casted as an unconstrained minimization problem. The effectiveness of this algorithm is demonstrated in a series of synthetic experiments and comparisons with the state-of-the-art spectral approaches.

Biography: Dimosthenis Pasadakis is Ph.D. candidate at Università della Svizzera italiana (USI) in Lugano, Switzerland, working at the Institute of Computing (CI) under the supervision of Olaf Schenk. The focus of his research is centered around algorithms for graph learning and combinatorial optimization for graph partitioning and clustering. Prior to that, he worked on fluid-structure interaction problems as part of his MSc thesis on Computational Science at USI, and studied Physics at the Aristotle University of Thessaloniki.