Expectation-maximization algorithm

This method will be to define a set of citation profiles and then self-consistently assign each case to the profile it best fits while at the same time adjusting the shape of the profiles to best fit the cases assigned to them. The means by which they accomplish this task is the expectation-maximization (EM) algorithm.

The EM algorithm is an established tool of statistics, but one that is relatively new to network analysis. This paper describe a different application to the analysis of the temporal profiles of citations.

In essence the EM algorithm is a method for fitting a model to observed data by likelihood maximization, but differs from the maximum likelihood methods most often encountered in the physics literature in that it does not rely upon Markov chain Monte Carlo sampling of model parameters. Instead, by judicious use of “hidden” variables, the maximization is performed analytically, resulting in a self-consistent solution for the best-fit parameters that can be evaluated using a relatively simple iteration scheme.

Reference

Expectation-maximization algorith on Wikipedia.

G.J. McLachlan, T. Krishnan, The EM Algorithm and Extensions. (Warning: this is an e version of this excellent book, I put it here just for introduction purpose, if you like this book, please borrow it in a library or buy it)