MC-Tree: Improving Bayesian Anytime Classification

In scientific databases large amounts of data are collected to create knowledge repositories for deriving new insights or planning further experiments. These databases can be used to train classifiers that later categorize new data tuples. However, the large amounts of data might yield a time consuming classification process, e.g. for nearest neighbors or kernel density estimators. Anytime classifiers bypass this drawback by being interruptible at any time while the quality of the result improves with higher time allowances. Interruptible classifiers are especially useful when newly arriving data has to be classified on demand, e.g. during a running experiment. A statistical approach to anytime classification has recently been proposed using Bayes classification on kernel density estimates.

In this paper we present a novel data structure called MC-Tree (Multi-Class Tree) that significantly improves Bayesian anytime classification. The tree stores a hierarchy of mixture densities that represent objects from several classes. Data transformations are used during tree construction to optimize the condition of the tree with respect to multiple classes. Anytime classification is achieved through novel query dependent model refinement approaches that take the entropy of the current mixture components into account. We show in experimental evaluation that the MC-Tree outperforms previous approaches in terms of anytime classification accuracy.

Authors: Kranen P., Günnemann S., Fries S., Seidl T.
Published in: Proc. of the 22nd International Conference on Scientific and Statistical Database Management (SSDBM 2010), Heidelberg, Germany, Springer LNCS
Publisher: Springer - Heidelberg, Germany
Sprache: EN
Jahr: 2010
Seiten: 252-269
ISBN: 978-3-642-13817-1
Konferenz: SSDBM
DOI: 10.1007/978-3-642-13818-8_19
Typ: Tagungsbeiträge
Forschungsgebiet: Data Analysis and Knowledge Extraction