BT* – An Advanced Algorithm for Anytime Classification

Abstract

In many scientific disciplines experimental data is generated at high rates resulting in a continuous stream of data. Data bases of previous measurements can be used to train classifiers that categorize newly incoming data. However, the large size of the training set can yield high classification times, e.g. for approaches that rely on nearest neighbors or kernel density estimation. Anytime algorithms circumvent this problem since they can be interrupted at will while their performance increases with additional computation time. Two important quality criteria for anytime classifiers are high accuracies for arbitrary time allowances and monotonic increase of the accuracy over time. The Bayes tree has been proposed as a naive Bayesian approach to anytime classification based on kernel density estimation. However, the employed decision process often results in an oscillating accuracy performance over time. In this paper we propose the BT* method and show in extensive experiments that it outperforms previous methods in both monotonicity and anytime accuracy and yields near perfect results on a wide range of domains.