Available under a Creative Commons Attribution Non-Commercial Share Alike 4.0 International Licence
Computer Sciences, Information Science
Concept drift is believed to be prevalent inmost data gathered from naturally occurring processes andthus warrants research by the machine learning community.There are a myriad of approaches to concept drift handlingwhich have been shown to handle concept drift with varyingdegrees of success.
However, most approaches make the keyassumption that the labelled data will be available at nolabelling cost shortly after classification, an assumption whichis often violated. The high labelling cost in many domainsprovides a strong motivation to reduce the number of labelledinstances required to handle concept drift. Explicit detectionapproaches that do not require labelled instances to detectconcept drift show great promise for achieving this.
Ourapproach Confidence Distribution Batch Detection (CDBD)provides a signal correlated to changes in concept without usinglabelled data. We also show how this signal combined with atrigger and a rebuild policy can maintain classifier accuracywhile using a limited amount of labelled data.
Lindstrom, P., MacNamee, B. & Delany, S. (2011) Drift Detection Using Uncertainty Distribution Divergence. 2nd International Workshop on Handling Concept Drift in Adaptive Information Systems (HaCDAIS) , Vancouver, Canada,11-14, December. doi:10.1109/ICDMW.2011.70