基于残差超网络的DNA微阵列数据分类

Using residual hypernetwork for the classification of DNA microarray data

DNA微阵列数据特征维度高,包含噪音,属性之间以及属性与样本类别之间有着复杂的关联性。然而传统超网络的超边一般是从训练集中随机选取属性而组成,难以保证超边质量,而且其分类性能受超边初始化过程影响很大,导致效果不稳定。针对传统超网络的这一局限,提出一种基于残差分析的超网络分类模型。残差算法根据显著性检验,首先假设属性相互独立,然后根据95%的置信水平,运用残差分析,用落入拒绝域的属性值对超网络的超边库进行初始化,以获取关联性较高的超边集合。然后采用梯度下降法进行超网络的演化学习。对急性白血病、前列腺癌和肺癌数据集的实验表明:与传统演化超网络分类器相比,该方法不仅有较高的分类精度,而且提高了分类的稳定性和收敛性。

DNA microarray datasets have the feature of high-dimension and are with noise. The association relationship among attributions and between attributions and sample sort is very complex. However,hyperedges of traditional hypernetwork generally consist of attributions which are randomly selected from a training dataset. Thus it is difficult to ensure the quality of hyperedges. And the classification performance of a traditional hypernetwork is seriously affected by the hyperedge initialization,so it is unstable. To conquer the defect of the traditional hypernetwork,a residual algorithm based hypernetwork model is proposed. The residual method is based on the significance test. First,we assume that the attributions are independent. And then according to the 95% confidence level and the residual analysis,the hyperedges library of the proposed hypernetwork is initialized with the attributions that fall into the refused domain. Thus the hyperedges are with strong relevance. Finally,the hypernetwork is evolved using a gradient descent algorithm. Compared with the traditional hypernetworks,the experimental results based on the acute leukemia,prostate and lung datasets show that the proposed scheme not only achieves better classification accuracy,but also improves the classification stability and the convergence ability of the hypernetwork model.