关键词: 数据流, 概念漂移, 主动学习, 自适应分类

Abstract:

Concept drift will cause the ability of data stream classification model to decrease with time，which requires the classification model with the ability of self?adaptation. However，most of the existing data stream classification models adapting to concept drift ignore the limited label of data stream. They usually assume that the coming data input to classification model can get the real label after the predicted label obtained. But，this assumption is unreasonable in some cases，as labeling data tends to be costly and time?consuming. In this paper，a new active learning method，ALBOD (Active Learning based on Boundary and Outlier Detection)，is proposed to solve the problem of the scarcity of data stream labels. Our method considers the problem of sampling bias which may occur in the process of active learning by combining the uncertainty active learning method with the BOD (Boundary and Outlier Detection) method. The first criterion is the most common active learning criterion，which ^{selects instances that are the most uncertain in terms of class membership. The latter curbs the sampling bias by using the fact that boundary samples and outliers can reflect the feature space. Compared to the 100% label algorithm OzaBagAdwin (OBA) and HoeffdingAdaptiveTree (HAT)，the proposed algorithm ALBOD can make the classification model maintain the same accuracy learning an average of about 20% labels under concept drift. The experiments show that though ALBOD is a sample combination of the above two criterions，it has a good active learning ability.}

Key words: data stream, concept drift, active learning, adaptive classification