Using Prediction Trends to Improve Concept Drift Tracking

M.M. Lazarescu (Australia)


Machine Learning, Concept Tracking, Knowledge Acquisition.


In this paper we describe a significantly revised version of the CWA on-line learning algorithm. Past methods devel oped to track drift only use previously observed data to de tect changes that occur in the concept, thus they are simply reacting to change. We present a method that not only uses previous knowledge to interpret the data but also to pre dict future data streams and identify possible changes in the concept. To track the drift, the algorithm also combines the concept change estimate with the prediction trend in or der to control the forgetting mechanism used to discard old data. The advantages of the new approach are that it helps the system converge faster to the target concept and it pro vides the system with a more effective way to interpret the data. We present the algorithm and analyse its performance using the STAGGER concept benchmark problem.

Important Links:

Go Back