基于混合采样和SE_ResNet_SVM的不平衡多分类研究

Unbalanced Multiclassification Study Based on Mixed Sampling and SE_ResNet_SVM

针对结构化多分类算法中不平衡数据集类别分布不均导致分类难度增加的问题,本文提出了一种基于混合采样、压缩与激励(squeeze and excitation, SE)模块、改进深度残差网络和支持向量机(support vector machines, SVM)的网络模型SNSMRS (SMOTEENNmixed residual networks-SVM network)。首先,通过合成少数过采样和编辑最近邻技术来改善数据分布;然后,构建融合SE模块与通过融合批次归一化和群组归一化的深度残差网络来提取特征;最后,通过SVM进行输出网络模型。其中,SE模块增强了模型对特征的区分能力,提升了模型的鲁棒性;基于融合归一化的残差网络受批次大小的影响较小,并且避免了传统神经网络梯度消失和精度退化等问题,增强了网络的稳定性与准确度;SVM可以根据特征向量在空间上的分布进行全部特征的分割,特征利用率高,提高了模型的分类精度。在7个不同规模和领域的非平衡公开数据集上进行了对比和消融实验,结果表明,本文所提的网络模型SNSMRS不仅优于其他深度学习模型,而且相对于未改良的ResNet,Macro-F1和G-mean值分别提升了约3%和4%,同时在4个数据集上的Macro-F1和G-mean值均超过了95%。

A network model SNSMRS(SMOTEENN-mixed residual networks-SVM network)based on hybrid sampling,squeeze and excitation(SE)module,improved deep residual network and support vector machines(SVM)is proposed to address the problem of uneven class distribution of unbalanced data sets in traditional structured multiclassification algorithms, which leads to increased classification difficulty. Firstly, the data distribution isimproved by synthesizing minority oversampling and editing nearest neighbors technique.Then the features are extracted by combining SE module and a deep residual network,improved with batch normalization and group normalization. Finally, the network modeluses support vector machine (SVM) to output the classification results. The SE moduleenhances the model’s feature differentiation ability and robustness. The improvements tothe ResNet, through fusion normalization, mitigate issues such as gradient vanishing andaccuracy degradation, and ensure stability and accuracy regardless of batch_size. Additionally,SVM enhances the classification accuracy by effectively utilizing feature vectors inspace to classify and extract features. Comparison and ablation experiments are conductedon seven unbalanced public datasets of various sizes and domains. The experimental resultsshow that the proposed model, SNSMRS, not only outperforms other deep learningmodels, but also increases the values of Macro-F1 and G-mean by approximately 3% and4%, respectively, compared with the original ResNet. Macro-F1 and G-mean values ofSNSMRS exceed 95% on four of the datasets, demonstrating its superior performance.