Bayesian Model Averaging of TAN Models for Clustering

Selecting a single model for clustering ignores the uncertainty left by finite data as to which is the correct model to describe the dataset. In fact, the fewer samples the dataset has, the higher the uncertainty is in model selection. In these cases, a Bayesian approach may be beneficial, but unfortunately this approach is usually computationally intractable and only approximations are feasible. For supervised classification problems, it has been demonstrated that model averaging calculations, under some restrictions, are feasible and efficient. In this paper, we extend the expectation model averaging (EMA) algorithm originally proposed in Santafe et al. (2006) to deal with model averaging of naive Bayes models for clustering. Thus, the extended algorithm, EMA-TAN, allows to perform an efficient approximation for a model averaging over the class of tree augmented naive Bayes (TAN) models for clustering. We also present some empirical results that show how the EMA algorithm based on TAN outperforms other clustering methods.