A Feature Fusion Approach for Multiple Signal Classification

This study proposes a feature fusion classifier (FFC) for multiple signal classification (MUSIC) by integrating different feature extraction methods and various machine learning algorithms. The functional principal component analysis (FPCA), common spatial pattern (CSP), discrete wavelet transform (DWT), and autoregressive (AR) model are considered to extract features from different domains. Using the features collected from the above methods, we propose to fit a logistic regression model with a model selection criterion to prevent overfitting and obtain relevant features at first. The relevant features are employed to learn different classifiers via different machine learning algorithms such as support vector machine, random forest, naive Bayes, extreme gradient boosting, and linear discriminant analysis. In addition, the technique of stacked generalization is also employed to integrate the classification results of these machine learning classifiers. Finally, the classifier with the smallest value of cross validation is selected to be the FFC. In our empirical study, three different MUSIC datasets are adopted to investigate the performance of the proposed method. Except for comparing the classification results of the FFC to the above machine learning classifiers, a fusion method that existed in the literature is also considered for comparison. The numerical results reveal that the FFC is capable of selecting useful features automatically and has robust and satisfactory performances for the 3 datasets.