一种适应概念漂移数据流的分类算法

针对带有概念漂移的数据流的分类问题,提出一种新颖的能够识别并且适应概念漂移数据流的分类算法。该算法将原始数据流沿着时间轴划分为若干数据块后,选择第一块中有代表性的数据作为样本训练模型,从而减轻了噪声和边界对分类精度的影响,使得漂移检测能较为全面且对离群点不过于敏感;此后对随后的数据块进行分类,并依据分类结果动态修正当前分类模型。实验结果表明:该方法能够根据数据流的当前状况自动调整分类模型,快速适应数据流概念漂移的情况,并得到较好的分类效果。

A self-adaptive classification method for concept-drifting data streams

A novel method was proposed for classifiying the concept-drifting data streams,which could track concept-drifting of data streams and quickly adapt to this change.After dividing a given data stream into several data blocks,it could choose the representative data from the first one for training model.The proposed method could alleviate the effects from noise and bordering data better,and be insensitive to outlier.Moreover,it used the created model for classifying each of the following data blocks,and used the classification results to dynamically adjust the current classification model.The experimental results showed that the proposed method could not only adjust classification model automatically according to the current status of data streams and quickly adapt to the situation of the concept drift,but also improve the classification performance.