一种基于分类器相似性集成的数据流分类研究

摘要

Classification of data streams has become one of hot research spots,and a new similarity-based dynamic ensemble algorithm was presented to deal with two critical problems,namely,concept drift and noise. Because adjacent data in data stream has the same concept with more probability, the new sub-classifier stands for the coming concept. Based on it,ensemble classifier is got by similarity weighted majority voting, and sub-classifier with worst performance is deleted for suiting for concept drift and noise. Experiment result on simulation data set shows the algorithm is best than other schema in classification accuracy and anti-noise.%数据流分类已成为当前研究热点之一,如何解决其中的概念漂移和噪声是关键问题,为此提出了一种新的基于分类器相似性的动态集成算法.由于数据流中相邻数据具有相同概念的概率较大,因此用最新基分类器代表数据流中即将出现的概念,同时基于此分类器求出基分类器之间的相似性作为权值进行加权多数投票,并根据相似性大小淘汰较弱基分类器以适应概念漂移和噪声.在标准仿真数据集上进行了仿真实验,结果表明该算法相比其他集成方法在抗噪性能和分类准确性方面均得到显著提高.