Abstract. We describe Athena: a system for creating, exploiting, and maintaining a hierarchy of textual documents through interactive miningbased operations. Requirements of any such system include speed and minimal end-user e ort. Athena satis es these requirements through linear-time classi cation and clustering engines which are applied interactively to speed the development of accurate models. Naive Bayes classi ers are recognized to be among the best for classifying text. We show that our specialization of the Naive Bayes classi er is considerably more accurate (7 to 29 % absolute increase in accuracy) than a standard implementation. Our enhancements include using Lidstone's law of succession instead of Laplace's law, under-weighting long documents, and over-weighting author and subject. We also present a new interactive clustering algorithm, C-Evolve, for topic discovery. C-Evolve rst nds highly accurate cluster digests (partial clusters), gets user feedback to merge and correct these digests, and then uses the classi cation algorithm to complete the partitioning of the data. By allowing this interactivity in the clustering process, C-Evolve achieves considerably higher clustering accuracy (10 to 20 % absolute increase in our experiments) than the popular K-Means and agglomerative clustering methods. 1

Given a set of possible models (e.g., Bayesian network structures) and a data sample, in the unsupervised model selection problem the task is to choose the most accurate model with respect to the domain joint probability distribution. In contrast to this, in supervised model selection it is a priori known that the chosen model will be used in the future for prediction tasks involving more \focused " predictive distributions. Although focused predictive distributions can be produced from the joint probability distribution by marginalization, in practice the best model in the unsupervised sense does not necessarily perform well in supervised domains. In particular, the standard marginal likelihood score is a criterion for the unsupervised task, and, although frequently used for supervised model selection also, does not perform well in such tasks. In this paper we study the performance of the marginal likelihood score empirically in supervised Bayesian network selection tasks by using a large number of publicly available classi cation data sets, and compare the results to those obtained by alternative model selection criteria, including empirical crossvalidation methods, an approximation of a supervised marginal likelihood measure, and a supervised version of Dawid's prequential (predictive sequential) principle. The results demonstrate that the marginal likelihood score does not perform well for supervised model selection, while the best results are obtained by using Dawid's prequential approach.