面向不平衡数据的多层神经网络模型

Multilayer neural network model for unbalanced data

传统的不平衡数据分类问题往往会因为类间数据不平衡造成分类器的性能下降。利用AUC(ROC曲线下的面积)为评价指标,结合单类F-score特征选择和遗传算法建立多层神经网络模型,选出对于不平衡数据分类更有利的特征子集,从而建立更适用于不平衡数据分类的深度模型。基于Tensor Flow建立多层神经网络模型,通过对4组不同UCI数据集进行测试,并与传统的机器学习算法如朴素贝叶斯、K最近邻、神经网络等进行对比验证。实验证明,所提模型在处理不平衡数据分类问题上的表现更优秀。

Classification of unbalanced data often has low performance of the classifier because of the unbalance of data between classes. Using AUC(the area under the ROC curve) as evaluation index, combined with one class F-score feature selection and genetic algorithm, a multilayer neural network model was established, and a more favorable feature set for unbalanced data classification was selected, so as to establish a deeper model suitable for classification of unbalanced data. Based on Tensor Flow, a multilayer neural network model was established. Using four different UCI datasets for testing, and comparing with the traditional machine learning algorithms such as Naive Bayesian, KNN, neural networks, etc, the performance of the proposed model built on the unbalanced data classification is more excellent.