Inference of Interaction Networks

Interaction networks are fundamental to our world, from trade interactions and social networks to the way epidemics spread. It is now becoming increasingly clear that the structure of temporal interaction networks crucially determines the dynamics of many important real-world processes. A very obvious example is the spread of diseases, but the same applies to other processes that we traditionally analyse without reference to interaction networks, such as opinion formation, collective decision making, and social learning.

Our own immediate application context is the organisation of animal societies, but the techniques we will develop are of a general nature and have far broader applications.

The fundamental aim of the project is to infer interaction networks from partial information.

Interaction networks can usually only be observed approximately, and this poses severe difficulties for analysing and predicting the processes that play out on them.

Most commonly interaction networks are approximated by contact networks derived from location tracking. There are two main reasons why this provides only partial information. Firstly, co-location creates the opportunity for an interaction, but it does not automatically mean that an interaction has occurred. Secondly, even contact networks themselves are normally incomplete. The most common method to construct contact networks is by location tracking, using video, GPS, or other sensors. Regardless of the method, data is rarely complete due to loss of contact, visual obscuring and other sensing problems.

The project will develop methods to infer the underlying interaction network based on partial information of the contact network and the observed spread of events. The fundamental approach is to define a network-based model of the spread of behaviour and then to use this model to infer the network structure as the most likely explanation of the observed spread of behaviours. Relevant methods include loopy belief propagation and Markov chain monte carlo. Test data will initially be generated with simulations in order to have a precisely defined ground truth, but we also aim to include experimental biological data.

The emphasis in the project will be on the development and evaluation of the computational methods, but interested students will optionally also have the opportunity to be involved in generating experimental data.