改进无监督极限学习机的不平衡数据分类

Improved Unsupervised Extreme Learning Machine Unbalanced Data Classification

针对传统的分类算法对不平衡数据分类的少数类数据分类准确率低的问题,本文基于模糊c均值聚类和SMOTE过采样技术,提出了改进的无监督极限学习机(FCM-US-ELM)。提出的方法通过模糊c均值聚类,将无标签的训练集正负类数据分为不同的簇,更新聚类中心和计算隶属度,然后按照规定的采样率在正类数据集上进行SMOTE过采样,使得训练集正、负类数据的个数趋于平衡,用新形成的训练集放入无监督极限学习机中训练。对比分析实验结果,提出的方法能够有效地减少数据的不平衡分布对正类数据分类正确率的干扰,得到更好的分类效果。在UCI数据集的分类实验中,新方法能够很好地处理数据的不平衡分类,达到了预期的效果。

To slove the problem that traditional classification algorithms have low classification accuracy for a few classes of data in unbalanced data sets,the paper proposes an improved unsupervised learning machine(FCM-US-ELM)based on Fuzzy c-means clustering and SMOTE undersampling technique.Firstly,the proposed method divides the positive and negative data of the training set into different clusters by fuzzy c-means clustering,then,updates the clustering center and calculates the membership degree Secondly,FCM-US-ELM performs SMOTE oversampling in the positive clusters at a specified sampling rate.Thirdly,SMOTE oversampling makes the number of positive and negative data in the training set tend to balance,then,the newly formed training set is put into the unsupervised extreme learning machine for training.Comparing the results of the test,the proposed method effectively reduces the interference of the unbalanced distribution of the data on the correct classification rate and obtains better classification results.In UCI data sets,the new method can handle the unbalanced classification of data well and achieve the desired results.