[Nets-seminars] Seminar, today 16:00 GS/102 66-72 Gower Street

[Richard G. Clegg]

I hope you will join us for today's talk -- it is by Dr. Damien Fay from 
the University of Bournemouth. The talk starts at 16:00 and is in GS/102 
in 66-72 Gower Street.

Applied Bayesian Compuation and Complex Network Models

The aim of this talk is to introduce a method called ABC which can be 
used to estimate parameters in models where the usual estimation 
techniques do not apply. In addition, we introduce ABC into the are of 
graph generator tuning and see that while it can be used there are 
several issues that need to be addressed.

This talk is broken into two sections. The first is somewhat a tutorial 
which focuses on ABC and gives an introduction to what we mean by 
complex processes and no clear means to estimate the parameters of that 
process (i.e. the blurb below will be explained in the talk). In the 
general area of model fitting, likelihood/Bayesian based approaches are 
popular not just because of the excellent parameter estimates but also 
because they return the posterior distribution of the parameter – i.e. 
we can see how probable it is that a parameter actually has a different 
value from the one we are using. One can think for example of the case 
where the posterior is almost uniform which essentially means the 
parameter estimate is useless; in this case care should be taken when 
deriving conclusions from the model.

Applied Bayesian Computation (ABC) is a method for estimating the 
posterior distribution for model parameters when the a likelihood based 
approach doesn't work; typically this occurs when the likelihood 
function a) doesn't exist, b) the likelihood is vanishingly small or c) 
it is just too complicated or time consuming to derive (especially if 
one is not convinced the model is appropriate in the first case). ABC 
has been around for many years but has recently become very popular due 
to some advances in the “summary statistic selection problem” - the 
Achilles heel of ABC.

The second half of the talk looks at graph topology generators. It may 
come as a surprise that 11 years after the introduction of the 
preferential attachment model there still exists no likelihood 
expression for that model. For the small world model one can only place 
bounds on the parameters. For models that include anything of greater 
complexity one is left without any option. This research thus asks the 
question, can ABC be applied to any graph topology generator? We then 
look at the answer and uncover some interesting technical issues leading 
to a set of heuristics for using ABC. We also take some real world data 
sets and show how using ABC we can select appropriate models for those 
datasets.

-- 
Richard G. Clegg,
Dept of Elec. Eng.,
University College London
http://www.richardclegg.org/