An adaptive ensemble classifier for mining complex noisy instances in data streams
Md. Rejaul Karim, Dewan Md. Farid · 2014
Real-time data streams classification is a challenging data mining task. In real-time streaming environments concepts of instances might change at any time such as weather predictions, astronomical and intrusion detection etc. To address this issue, we present an adaptive ensemble classifier for data streams classification, which uses a set of decision trees for mining complex noisy instances in data streams. The ensemble model updates automatically so that it represents the most recent concepts in data streams. In each iteration, the ensemble model generates a new training data from original training dataset, then builds a decision tree using new training data and assigns a weight to the tree based on its classification accuracy on original training instances. Also it updates the weight of training instances in training dataset. We tested the performance of the proposed ensemble classifier against that of existing C4.5 decision tree classifier using real benchmark datasets from UCI (University of California, Irvine) machine learning repository. The experimental results prove that the proposed ensemble classifier shows great flexibility and robustness in data streams classification.