Boosting Extensions to Find Anomaly Structure in Data

• Activity Number: 59
• Type: Invited
• Date/Time: Sunday, August 8, 2004 : 6:00 PM to 7:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Virginia L. Wheway*+
• Companies: Boeing Company
• Keywords: data-mining ; noisy ; boosting ; machine learning ; classification ; cluster

Abstract

Recent advances in data mining have led to the development of a method called "boosting." Instead of building a single model, boosting sees several models being built using weighted versions of the original data. These models are then combined into a single prediction model via voting. Studies have demonstrated that boosting leads to significantly lower prediction error on unseen data. This poster demonstrates how the method of boosting may be extended beyond its original aims of improved prediction. Simple plots of specific boosting statistics may be used as tools to detect noisy data and unearth structure within datasets. Whether or not this "suspect" data occurs in groups or as single observations may also be determined. An industrial dataset will be used to demonstrate the process, and show the power of detecting unknown clusters within datasets. Proposed extensions for this research include testing for the threshold of cluster size able to be detected. Research into the method's extension to continuous data would certainly have many opportunities for existing datasets in many sectors, particularly for data containing rare events and data usually considered to be noisy.