一种基于聚类提升的不平衡数据分类算法

摘要

Imbalanced data exist widely in the real world and their classiifcation is a hot topic in the ifeld of machine learning. A clustering-based enhanced AdaBoost algorithm was proposed to improve the poor classiifcation performance produced by the traditional algorithm in classifying the minority class of imbalanced datasets. The algorithm firstly constructs balanced training sets by the clustering-based undersampling, using K-means clustering to cluster the majority class and extract cluster centroids and then merge with all minority class instances to generate a new balanced training set. To avoid the declining of the classiifcation accuracy caused by the shortage of training sets owing to too few minority class samples, SMOTE (Synthetic Minority Oversampling Technique) combining the clustering-based undersampling was used. Next, the misclassiifcation loss function in the basic classiifer of the AdaBoost algorithm was modiifed based on the cost-sensitive learning theory to assign asymmetric misclassiifcation losses to samples of different classes. The experimental results show that, the proposed algorithm makes the model training samples more representative and greatly increases the classiifcation accuracy of the minority class, keeping the overall classiifcation performance.%不平衡数据分类是机器学习研究领域中的一个热点问题。针对传统分类算法处理不平衡数据的少数类识别率过低问题，文章提出了一种基于聚类的改进AdaBoost分类算法。算法首先进行基于聚类的欠采样，在多数类样本上进行K均值聚类，之后提取聚类质心，与少数类样本数目一致的聚类质心和所有少数类样本组成新的平衡训练集。为了避免少数类样本数量过少而使训练集过小导致分类精度下降，采用少数过采样技术过采样结合聚类欠采样。然后，借鉴代价敏感学习思想，对AdaBoost算法的基分类器分类误差函数进行改进，赋予不同类别样本非对称错分损失。实验结果表明，算法使模型训练样本具有较高的代表性，在保证总体分类性能的同时提高了少数类的分类精度。