一种面向动态不平衡数据流的集成超限学习机分类算法

An Ensemble Classification Method of Extreme Learning Machine for Dynamic Imbalanced Data Steams

随着数据收集、存储和传输技术的快速发展,数据流的挖掘处理技术正在成为机器学习中的一个热点问题。在许多情形下,持续到达的数据之间可能会呈现出不平衡的态势,甚至是动态不平衡,这给许多机器学习算法造成了困难。文中提出面向动态不平衡数据流的集成超限学习机算法,设计了数据流中不平衡率变化的快速监测方法,修正了历史数据不平衡率的计算方式,使其更接近不平衡率的实时变化,并结合超限学习机的特点,将增量学习与集成学习结合。定期剔除权重低的基分类器,利用新到达的数据更新集成中的基分类器和训练新的基分类器。该方法针对动态不平衡数据流设计,具有很好的学习能力,同时也能适用于静态或者平衡的数据流的分类。实验中,将该方法与其他几种常用的方法在一些不同类型的数据流上进行了比较,结果表明,文中方法的分类性能更好。

With the rapid development of data collection,storage and transmission technology,how to mine data stream effectively is becoming a hot issue in machine learning.In many cases,the continuously arriving data may appear imbalanced,even dynamic imbalanced,which bring difficulties to many machine learning algorithms.In this paper,an Ensemble framework of Extreme Learning Machine for Dynamic Imbalanced data stream(EELMDI)was presented.We have developed a fast detection method for the change of imbalance rate of data stream,and modified the calculation method of historical data􀆳s imbalance rate to make it closer to the real-time change of imbalance rate of data stream.Dynamic weight of each base classifier is determined based on its learning performance for data stream and the characteristics of the extreme learning machine.The method removes the learners based on their classification performance,and it updates base classifiers in the ensemble and trains new base classifier with new arrival data at regular intervals.This method is designed for dynamic imbalanced data stream and has good learning performance.It can also be applied to stationary or balanced data stream.In the experiment,we compare the proposed method with other state-of-the-art algorithms on several different forms of data streams.The results show that the proposed method achieves the better performance.