Functional connectivity inference from time-series of events and application to computer networks

Messager, Antoine (2020) Functional connectivity inference from time-series of events and application to computer networks. Doctoral thesis (PhD), University of Sussex.

[img] PDF - Published Version
Download (8MB)


Today’s commercial and Internet Service Provider’s networks are large, heterogeneous, fast-evolving and commonly emit millions of events per second. Monitoring, responding to, and predicting incidents is a key challenge for network management operators. This thesis argues that the concept of functional connectivity, widely used in neuroscience and defined in terms of the presence of statistical dependencies between nodes, can unlock informational value about event logs and assist operators in identifying (and even predicting) service outages. However, existing functional connectivity inference methods are not adapted to event data from computer networks. These methods may either require unavailable models of event propagation, be computationally too costly for large networks and/or long recordings, be not adapted to sparse and discrete activity, and/or assume a static network topology. We thus first describe in depth a major commercial network to highlight the challenges faced by network operators and the opportunities offered by thinking in terms of functional connectivities. Next, using a pair of independent Bernoulli processes as reference, we develop a new statistic aimed at measuring coupling strength by quantifying deviation from independence. However, because many statistics will be large by chance, identifying functional edges from the distribution of statistics over every pair of nodes is challenging. Hence, we then develop a method that infers, in specific contexts, the function that associates each statistic to the probability it accounts for the presence of a functional edge. Next, using the previously described statistic, we propose a novel framework to infer a time-varying functional topology from the time-series of emitted events. Applying this paradigm to two major commercial networks, we show that it can reveal a priori unknown groups of devices providing particular services. Finally, we argue that this thesis has implications beyond assisting network monitoring, such as enabling more robust inference of functional connectivity in neuroscience.

Item Type: Thesis (Doctoral)
Schools and Departments: School of Engineering and Informatics > Informatics
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5101 Telecommunication > TK5105.5 Computer networks
Depositing User: Library Cataloguing
Date Deposited: 29 Apr 2020 15:21
Last Modified: 29 Apr 2020 15:21

View download statistics for this item

📧 Request an update