基于聚类假设的数据流分类算法

发布时间：2018-03-09 09:07

本文选题：概念漂移　切入点：数据流　出处：《模式识别与人工智能》2017年01期 　论文类型：期刊论文

【摘要】：获取数据流上样本的真实类别的代价很高,因此标记所有样本的方式缺乏实用性,而随机标记部分样本又会导致模型的不稳定.针对上述问题,文中提出基于聚类假设的数据流分类算法.基于通过聚类算法分到同类中的样本可能具有相同类别这一聚类假设,利用训练数据集上的聚类结果拟合样本的分布情况,在分类阶段有目的性地选取很难分类或潜在概念漂移的样本更新模型.为了训练数据集上每个类别的样本,建立各自对应的基础分类器,当数据流中样本的类别消失或重现时,只需要冻结或激活与之对应的基础分类器,而无需再重新学习之前已经掌握的知识.实验表明,文中算法能够在适应概念漂移的前提下,减少更新模型需要的样本数量,并且取得和当前数据流上的分类算法相当或更好的分类效果.
[Abstract]:The cost of obtaining the true categories of samples on a data stream is very high, so the way all samples are labeled is not practical, and random marking of parts of samples can lead to instability of the model. In this paper, a data stream classification algorithm based on clustering hypothesis is proposed. Based on the clustering assumption that the samples in the same class may have the same class, the clustering results on the training data set are used to fit the distribution of the samples. In order to train the samples of each category in the data set, a corresponding basic classifier is established, when the class of the sample in the data stream disappears or reappears. It is only necessary to freeze or activate the corresponding basic classifier without relearning the previous knowledge. Experiments show that the proposed algorithm can reduce the number of samples needed to update the model on the premise of adapting to the drift of the concept. And the classification algorithm is comparable to or better than the current classification algorithm on the data stream.
【作者单位】：福建农林大学计算机与信息学院;
【基金】：福建省自然科学基金项目(No.2016J01280)资助~~
【分类号】：TP311.13

【参考文献】