Unsupervised Ensemble Feature Selection for Underwater Acoustic Target Recognition

Abstract: The problems of feature redundancy and label information lacking of underwater acoustic target data and potential local optima of general feature selection algorithms result in poor classification performance and stability in the process of the underwater acoustic target recognition. The unsupervised feature selection using feature similarity can effectively remove redundant features without label information, and the ensemble method can improve the generalization accuracy and stability. Therefore, we proposed an algorithm of unsupervised ensemble feature selection using feature similarity (UEFSUFS) for underwater acoustic target recognition, which primarily generates several sets of results of the feature selection algorithm using diverse training sets by sampling from original training set with replacement, subsequently aggregates those results using voting method. The SVM classification performance and stability of the proposed method are examined on the UCI Sonar dataset and a real-world underwater acoustic target dataset. Experimental results on the two datasets show that the proposed method can improve the average classification rate of SVM and the Jccard index, which indicates that the proposed method can remove redundant features and improve stability of feature selection, thereby leading to better classification performance.Keywords: underwater acoustic target recognition; unsupervised; ensemble; feature selection.