基于特征选择和聚类的动态选择性集成模型

A Dynamic Ensemble Selection Model Based on Feature Selection and Clustering

关键词:  特征选择  动态选择性集成  支持向量机  蚁群优化算法  辐射源个体识别  二分类问题

Key words:  feature selection  dynamic ensemble selection  support vector machine  ant colony Optimization  specific emitter identification  binary classification

摘要:

为提高辐射源个体识别的准确率,降低动态选择性集成的计算复杂度，本文提出基于特征选择和聚类的动态选择性集成模型(FSC-DES)。利用归一化皮尔森相关系数法度量不同基分类器间混淆矩阵的差异性，以各基分类器准确率最高及基分类器间差异性最大为目标，得到基分类器集合和对应特征子集集合。利用聚类方法将验证集划分为若干类，以验证集分类准确率最高为目标，为每簇验证集选择最优的基分类器子集和对应的特征子集。在测试阶段，对测试集进行聚类，仅比较每簇测试样本和每簇验证样本数据分布的最大均值差异值，减少运算时间。每簇测试样本在相似度最高的验证集所对应的特征子集集合和基分类器子集下进行预测，并根据不同权重基分类器预测结果的加权和进行最终决策。为验证方法的必要性和优越性，将本文方法与传统集成学习方法进行对比，结果表明，本文方法在信噪比分别为10、5 dB的条件下，分类准确率均提升约5%，具有更好的分类效果和泛化性能。

Abstract:

To improve the accuracy of emitter individual recognition and reduce the computational complexity of dynamic selective ensemble,a Dynamic Ensemble Selection model based on Feature Selection and Clustering (FSC-DES) is proposed.The normalized Pearson correlation coefficient method is used to measure the difference of the confusion matrix between base classifiers,and the base classifier set and the corresponding feature subset set are obtained with the goal of maximizing the accuracy of each base classifier and the difference between base classifiers.The validation set is divided into several classes by the clustering method,and the optimal base classifier subset and corresponding feature subset are selected for each cluster validation set with the goal of maximizing classification accuracy of the validation set.In the testing phase,clustering is performed for the test set,and only the maximum mean difference of the data distribution is compared between each cluster of test samples and each cluster of validation samples to reduce the operation time.Each cluster of test samples is predicted under the feature subset set and base classifier subset corresponding to the validation set with the highest similarity,and the final decision is made according to the weighted sum of the prediction results of different weight base classifiers.Furthermore,the method proposed in this paper is compared with the conventional integrated learning method to assess the necessity and superiority of the method.The results show that the method proposed in this paper has the improvement of about 5% in the classification accuracy when the signal-to-noise ratio is 10 dB and 5 dB,respectively,demonstrating better classification effect and generalization performance.