Copula Network Classifiers (CNCs)
Gal Elidan ; JMLR W&CP 22: 346-354, 2012.
The task of classification is of paramount importance and extensive research has been aimed at developing general purpose classifiers that can be used effectively in a variety of domains. Network-based classifiers, such as the tree augmented naive Bayes model, are appealing since they are easily interpretable, can naturally handle missing data, and are often quite effective. Yet, for complex domains with continuous explanatory variables, practical performance is often sub-optimal. To overcome this limitation, we introduce Copula Network Classifiers (CNCs), a model that combines the flexibility of a graph based representation with the modeling power of copulas. As we demonstrate on ten varied continuous real-life datasets, CNCs offer better overall performance than linear and non-linear standard generative models, as well as discriminative RBF and polynomial kernel SVMs. In addition, since no parameter tuning is required, CNCs can be trained dramatically faster than SVMs.