Semisupervised learning in pattern recognition with concept drift

Abstract

We describe the use of semisupervised learning to adapt decision rule in the pattern recognition problem with concept drift. There has been generated a new classifier using additional dataset that becomes available from the changing environment. The classifier is a combined cluster structure with a modified weighted Bayesian decision rule, where the weights are dynamically updated using the classifier’s current decision. The probability density functions are identified in each cluster and the decision rule is defined as a distribution mixture. The adaptation mechanism allows the algorithm to track the environment changes by weighting the most recent and relevant cluster higher. The adaptive algorithm is described, and its performance is compared to the static one by using specific model problem.