New Generation Computing, 21(2003)149-161
Ohmsha, Ltd. and Springer-Verlag

Chance Discovery and Learning Minority Classes

Tu Bao HO and Duc Dung NGUYEN
Japan Advanced Institute of Science and Technology
Tatsunokuchi, Ishikawa, 923-1292 Japan
{bao,dungduc}@jaist.ac.jp

Received 12 March 2001

Abstract

Chances are viewed in chance discovery as events/situations with significant impact on human decision making. In this research context we are particularly interested in a subset of chances that are unexpected or contradictory with human common knowledge, and the human role that we consider as an essential factor in finding such chances. We first introduce the method LUPC that can learn minority classes from large unbalanced datasets. With its visualization tools as well its exclusive and inclusive constraints, LUPC allows the user to actively participate in and to incorporate background knowledge in the chance discovery process. We then present case studies in which LUPC is used to support the user in discovering significant unexpected chances from stomach cancer and hepatitis databases.

Keywords: Learning Minority Class, Chance Discovery, Visualization, Domain Knowledge, Medical Data.

[Back]