Inferring Useful Heuristics from the Dynamics of Iterative Relational Classifiers

In this paper we consider dynamical properties of simple iterative relational classifiers. We conjecture that for a class of algorithms that use label-propagation the iterative procedure can lead to nontrivial dynamics in the number of newly classified instances. The underlaying reason for this nontriviality is that in relational networks true class labels are likely to propagate faster than false ones. We suggest that this phenomenon, which we call two-tiered dynamics for binary classifiers, can be used for establishing a self-consistent classification threshold and a criterion for stopping iteration. We demonstrate this effect for two unrelated binary classification problems using a variation of a iterative relational neighbor classifier. We also study analytically the dynamical properties of the suggested classifier, and compare its results to the numerical experiments on synthetic data.