卷积神经网络与人工水母搜索的图特征选择方法

Convolutional Neural Network and Artificial Jellyfish Search-based graph feature selection method

目前,卷积神经网络(Convolutional Neural Network,CNN)模型在处理图像数据时分类效果较差,人工水母搜索(Artificial Jellyfish Search,AJS)算法收敛速度慢,迭代次数多,特征选择的效果不理想.针对上述问题,提出一种基于CNN和AJS的图特征选择方法 .首先,使用CNN来提取特征,将生成的特征图进行图嵌入降维,再使用AJS算法进行特征选择,把得到的特征输入分类器,进行模型训练和评估;然后,在图嵌入阶段,将特征图进行随机游走,并通过添加特征权重计算节点的游走概率来增强权重大的节点的游走概率,提高分类精度;最后,在AJS算法中引入余弦公式对人工水母的位置进行更新,充分考虑特征向量之间的相似性,提高算法的收敛速度并减少迭代次数.在10个基准函数上进行实验,结果表明改进的AJS算法具有较好的优化性能.在四个数据集上,将提出的算法与其他算法进行对比实验,实验结果表明,提出的算法能提高分类精度,减少迭代次数.

At present,the classification effect of Convolutional Neural Network(CNN)model is poor when dealing with image data,and the Artificial Jellyfish Search(AJS)algorithm has slow convergence speed and large number of iterations,which leads to unsatisfactory feature selection effect.This paper proposes a graph feature selection method based on CNN and AJS.Firstly,the CNN extracts features,and the generated feature map is used for graph embedding dimensionality reduction.The AJS algorithm is used for feature selection,and the obtained features are input into the classifier for model training and evaluation.Then,in the graph embedding stage,the feature map is randomly walked and the walk probability of the node is calculated by adding the feature weight,which enhances the walk probability of the node with significant weight and improves the classification accuracy.Finally,the cosine formula is introduced into the AJS algorithm to update the position of the artificial jellyfish,and the similarity between the feature vectors is fully considered to improve the convergence speed and reduce the number of iterations.Experimental results on ten benchmark functions show that the improved AJS algorithm has better optimization performance.Compared with other algorithms on four datasets,the proposed algorithm improves the classification accuracy and reduces the number of iterations.