基于弹性网回归的水下目标无监督特征选择算法

Unsupervised Feature Selection Algorithm Based on Elastic-net Regression for Underwater Target

摘要:

针对水声数据的特征中含有大量冗余、不相关和噪声特征，导致水声目标识别正确率降低的问题，提出了一种新的水下目标特征选择方法——基于弹性网回归的无监督特征选择算法(Unsupervised fea- ture selection algorithm based on elastic-net regression，UFSER)。 该算法利用谱回归得到高维水声数据和其低维表示之间的回归系数矩阵，并且在回归框架中加入弹性网惩罚项优化求解回归系数矩阵；最后，对回归系数矩阵进行稀疏化从而对特征的分类性能进行评价。 使用实测水声数据集和 UCI 声呐数据集进行特征选择和 SVM 分类实验。 实验结果表明：在特征数目分别减少 60.6%和 60%的情况下，分类识别正确率较特征选择前提升了 1.05%和 6.6%。

Abstract:

There are many redundant， irrelevant and noise features in underwater acoustic data feature， which reduce the recognition accuracy of underwater acoustic target． To solve this problem， this paper proposes a new feature selection algorithm for underwater target recognition—unsupervised feature selection algorithm based on elastic-net regression (UFSER)． The proposed al? gorithm uses spectral regression to calculate a regression coefficient matrix between high-dimensional original underwater acoustic data and its low-dimensional representation． In the regression framework， an elastic-net penalty term is applied to optimize the re? gression coefficient matrix； finally， the classification performance of features is evaluated with the regression coefficient matrix spar? sification． In this paper， feature selection and classification experiments are performed by using the measured underwater acoustic dataset and the UCI sonar dataset． The experimental results show that the number of features is reduced by 60.6% and 60%， while the classification identification accuracy is improved by 1.05% and 6.6% compared with the accuracy before feature selection．