Latent classification models

One of the simplest, and yet most consistently well-performing setof classifiers is the \NB models. These models rely on twoassumptions: $(i)$ All the attributes used to describe an instanceare conditionally independent given the class of that instance,and $(ii)$ all attributes follow a specific parametric family ofdistributions.  In this paper we propose a new set of models forclassification in continuous domains, termed latent classificationmodels. The latent classification model can roughly be seen ascombining the \NB model with a mixture of factor analyzers,thereby relaxing the assumptions of the \NB classifier. In theproposed model the continuous attributes are described by amixture of multivariate Gaussians, where the conditionaldependencies among the attributes are encoded using latentvariables. We present algorithms for learning both the parametersand the structure of a latent classification model, and wedemonstrate empirically that the accuracy of the proposed model issignificantly higher than the accuracy of other probabilisticclassifiers.