基于差分孪生卷积神经网络的大规模不平衡数据分类算法  

CLASSIFICATION ALGORITHM OF BIG SCALE IMBALANCED DATA BASED ON THE DIFFERENTIAL SIAMESE CONVOLUTION NEURAL NETWORKS

作　　者：任佳丽[1] 王文晶 Ren Jiali;Wang Wenjing(Department of Information Engineering,Shanxi Vocational and Technical College,Taiyuan 030031,Shanxi,China;College of Information,Business College of Shanxi University,Taiyuan 030031,Shanxi,China)

机构地区：[1]山西交通职业技术学院信息工程系,山西太原030031 [2]山西大学商务学院信息学院,山西太原030031

出　　处：《计算机应用与软件》2019年第11期267-274,共8页Computer Applications and Software

基　　金：山西省高校科技创新项目(2015107)

摘　　要：传统基于支持向量机的不平衡数据分类算法包含矩阵运算,无法应用于大规模的不平衡数据集.针对这种情况,提出基于差分孪生卷积神经网络的大规模不平衡数据分类算法.设计差分卷积机制增强卷积神经网络的深度结构表示能力,在不改变滤波器数量的情况下提高模型的判别能力.通过差分孪生卷积神经网络分别优化每个类的特征图,每个类关联多个超平面,根据输入样本与超平面的距离决定输出样本的类标签.基于多组不平衡数据集的实验结果表明,该算法实现了较好的分类性能.The traditional classification algorithms for imbalanced data based on the support vector machine include matrix operations,which cannot be applied in big scale imbalanced datasets.In view of this,we propose a classification algorithm for big scale imbalanced data based on the differential Siamese convolution neural networks.We designed a differential convolution mechanism to enhance the deep structure representation ability of convolution neural networks,so that it can improve the discriminative ability of models without changing the number of filters.The Siamese mechanism was introduced to convolution neural networks.The feature maps of each class were optimized by differential convolution Siamese networks.Each class was associated with several hyperplanes,and the model decided the class labels of output samples according to the distances between the input samples and hyperplanes.Experimental results based on several imbalanced datasets show that the proposed algorithm performs better classification results,and has faster speed.

关 键 词：深度学习 数据分类 不平衡数据集 卷积神经网络 深度神经网络 孪生神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]