Estimação de grafos não direcionais aplicada à aprendizagem de máquina.

Abstract

Dynamical systems may be modelled as undirected graphs in which the nodes represent the devices of the system and the edges represent possible connections amongst them. After the rapid development of heterogeneous networks, which provided a massive interconnection for different types of devices, the analysis of data outputed from these devices has become a paramount task to understand the structure underpinning such dynamical systems. In this way, this work proposes an estimator for the Laplacian matrix of undirected graphs in order to infer the structure and thus the interconnections (correlations) amongst the devices. We have used the Bayesian approach in order to take into account the spectral properties of the Laplacian matrix into the project of the estimator. The proposed estimator was applied in data clustering, which is a subfield of unsupervised machine learning. The experimental results indicate that the estimator’s performance is satisfactory, and it is expected when compared with results available in the literature