基于非平衡问题的卷积神经网络分类模型

A CONVOLUTION NEURAL NETWORK CLASSIFICATION MODEL BASED ON IMBALANCED PROBLEM

由于数据分布的不平衡,传统的分类模型常常会受多数类的影响而降低分类准确率。因此为提升对非平衡数据的分类性能,提出新的卷积神经网络分类模型CNN-EMWRA-WCELF,其中EMWRA(Expectation Maximization Weighted Resampling Algorithm)是对EM算法的优化。融合高斯混合模型与采样算法,对初始数据集进行精确采样,以此降低训练数据集的非平衡度。另外,所提的WCELF(Weighted Cross Entropy Loss Function)函数,可根据样本权重对少数类和多数类赋予不同的代价损失。最后,以F1和G-mean为评价指标,将CNN-EMWRA-WCELF模型在竞赛数据集上与其他分类模型相比较,结果均为第一,表明其能够很好地提高少数类分类的正确率。

Due to the imbalance distribution of data,the traditional classification model is often affected by most classes and reduces the classification accuracy.Therefore,in order to improve the classification performance of imbalanced data,this paper proposes a new convolutional neural network classification model CNN-EMWRA-WCELF,in which EMWRA(expectation maximization weighted resampling algorithm)is the optimization of EMalgorithm.It combined Gaussian mixture model and sampling algorithm to accurately sample the initial data set,so as to reduce the imbalanced degree of the training data set.In addition,the proposed WCELF(weighted cross entropy loss function)could assign different cost losses to minority and majority classes according to the sample weight.Taking F1 and G-mean as evaluation indexes,the proposed CNN-EMWRA-WCELF model was compared with other classification models on competition data sets.The results show that theevaluation indexes of the proposed model are best,indicating that it can improve the accuracy of minority class classification.