Jannik Thümmel: Inferring network topology from complex contagion dynamics

BCCN Berlin / Technische Universität Berlin

 

Abstract

The central question that motivates this study is which principles underlie the trans-
fer of information in collective behaviour. To progress towards understanding these
principles, we study networks of interaction that govern the pathways along which
information can travel.

We focus on two types of local interactions based on a metric and a k-nearest
neighbour rule respectively. The topology of the resulting networks induces different
patterns of activity spreading through the system. The activity is modeled as a
complex contagion process. Out aim is to differentiate between the topologies based
only on the observable activity.

We explore two methodological approaches to this inference problem. Inspired
by the exciting developments in the field of machine learning we focus on data-
driven models that can extract useful patterns. The first method we propose uses a
logistic regression model to understand the statistics of local interactions. The sec-
ond method draws on the strength of convolutional neural networks to find relevant
patterns of interaction from a global view of the entire system.

Both methods studied in this work are shown to yield promising results. We
take steps towards understanding the limitations of these approaches and discuss
possible ways in which we could extend them to real-world systems.

 

Additional Information

Master Thesis Defense

 

Organized by

Dr. Pawel Romanczuk  & Prof. Dr. Henning Sprekeler  / Lisa Velenosi