基于改进BP神经网络的非均衡数据分类算法

传统的分类算法大都默认所有类别的分类代价一致，导致样本数据非均衡时产生分类性能急剧下降的问题.对于非均衡数据分类问题，结合神经网络与降噪自编码器，提出一种改进的神经网络实现非均衡数据分类算法，在神经网络模型输入层与隐层之间加入一层特征受损层，致使部分冗余特征值丢失，降低数据集的不平衡度，训练模型得到最优参数后进行特征分类得到结果.选取UCI标准数据集的3组非均衡数据集进行实验，结果表明采用该算法对小数据集的分类精度有明显改善，但是数据集较大时，分类效果低于某些分类器.该算法的整体分类效果要优于其他分类器.

Unbalanced Data Classification Algorithm Based on Improved BP Neural Network

Most of the traditional classifications algorithms have the same classification cost of all categories, which results in a sharp decline in classification performance when the sample data are unbalanced. As to the problem of unbalanced data classification, we combine neural network with denoising auto-encoder and put forward a kind of improved neural network to realize unbalanced data classification algorithm. The algorithm adds a layer called feature damaged layer between input layer and hidden layer. Thus some redundant feature values are lost, and the unbalance degree of data set is reduced. And the results can be obtained after training model obtains optimal parameters and deals with the classification based on feature. It selects three sets of UCI standard unbalanced data sets for experiment. The results show that the accuracy of the algorithm for small data set classification is improved obviously, but when the data set is larger, the classification effect is lower than some classifier. And the overall classification performance of the proposed algorithm is better than other classifiers.