Iterative relational classification through three-state epidemic dynamics

Relational classification in networked data plays an important role in many problems such as text categorization, classification of web pages, group finding in peer networks, etc. We have previously demonstrated that for a class of label propagating algorithms the underlying dynamics can be modeled as a two-state epidemic process on heterogeneous networks, where infected nodes correspond to classified data instances. We have also suggested a binary classification algorithm that utilizes non-trivial characteristics of epidemic dynamics. In this paper we extend our previous work by considering a three-state epidemic model for label propagation. Specifically, we introduce a new, intermediate state that corresponds to "susceptible" data instances. The utility of the added state is that it allows to control the rates of epidemic spreading, hence making the algorithm more flexible. We show empirically that this extension improves significantly the performance of the algorithm. In particular, we demonstrate that the new algorithm achieves good classification accuracy even for relatively large overlap across the classes.