Identifying cause-effect relationships in time-series data

Master Thesis

 

Identifying cause-effect relationships in time-series data

Background

Agama collects vast amounts of time-series data that details the performance of a video delivery network. For a user of Agama’s solutions (typically a technician at an operations center) some of the interesting things that the user would like to see / questions to answer are:

  • How many errors in total have occurred in the system now, in the last 15 min, in the last hour, etc.
  • Which services are responsible for the error count, now and over time?
  • Are errors more likely to occur on a certain device than on others?

While such questions are possible to identify by looking closer at the data, automatically identifying the root causes of such errors can be extremely valuable and save time in performing remediation actions.

Assignment

At Agama, we have built a simulator that can simulate errors in delivery networks and reproduce the expected client-side effects for a large subset of these errors. The problem to solve in this thesis is to build models that will point to one or more of the most likely causes of the problems.

For example, consider that a 5% packet loss rate in a multicast switch in the IPTV part of the TV network. When we introduce such an error in our simulator, all the IPTV clients connected to the faulty delivery node will experience playback issues, and generate appropriate numbers for freezes, qoeErrors, etc. We can now make a causal model of how a 5% packet loss will appear by looking at the simulated output data. By correlating this information with the data observed at IPTV switches, we can identify a potentially faulty switch.

The goal of this thesis is to exhaustively execute such network- and device-related errors, collect  the expected client-side behavior and model them in such a way that is the most likely cause (this can also be probabilistic) for the error.

Models based on decision-trees/forests would be the initial approach, as they can represent the causal nature of errors based on the different observed parameters.

Apply

To learn more about this master thesis or to apply, contact:

Aner Gusic
aner.gusic@agama.tv

To learn more about this master thesis or to apply, contact:

Aner Gusic
aner.gusic@agama.tv

Want to know more?

Would you like to know more about Agama and our story?