Concept Drift Detector Selection for Hoeffding Adaptive Trees

摘要

Dealing with evolving data requires strategies for detecting and quantifying change, and forgetting irrelevant examples during the model revision process. To design an adaptive classifier that is suitable for different types of streams requires us to understand the characteristics of the data stream. Current adaptive classifiers have built-in concept drift detectors used as an estimator at each node. Our research aim is to investigate the usage of different drift detectors for Hoeffding Adaptive Tree (HAT), an adaptive classifier. We proposed three variants of the proposed classifier, called HAT_(SEED), HAT_(HDDMA), and HAT_(PHT).