Edward Ott

University of Maryland

Using Machine Learning and Causality-Destroyed Surrogate Data to Infer Network Links from Nodal Time Series
5:35pm, Room St Clair 3B

We consider the general problem of network inference from nodal time series data, introducing two new tools: (i) a machine learning technique for obtaining numerical scores from nodal time series data where the scores reflect the likelihood of links between node pairs, and (ii) a new type of surrogate data serving as a basis for statistical analysis yielding link information from the score set for commonly occurring cases where the distribution of scores would otherwise make link inference problematic. We assess these new tools by application to three different data sets --- numerically generated data from a network of Lorenz oscillators [1], from both experimental data and and numerical modeling of a network of interacting lasers with delays along there network links [2], and from publicly available neuronal data from C. elegans. We compare the performance of our machine learning technique for obtaining scores with the transfer entropy method.

[1] A. Banerjee, J. Pathak, R. Roy, J.G. Restrepo, E. Ott, CHAOS 29, 121104 (2019).
[2] A. Banerjee, J.D. Hart, R. Roy, E.Ott, Phys. Rev. X 11, 031014 (2121).