Adaptive Personalized Feature Selection for underwater Acoustic Signals

To improve the classification performance of underwater acoustic (UWA) signals, features from the time domain, frequency domain, and time-frequency domain are extracted from these signals. However, this extraction process introduces irrelevant information for individual categories. To address this issue, a novel adaptive personalized feature selection method is proposed to find all relevant features for specific categories within a UWA signal dataset. Utilizing Tree SHAP (Shapley Additive Explanations), the importance value of each feature is calculated based on its contribution to the prediction, considering all possible combinations of features. According to these SHAP values, a time-shifted personalized model is developed to dynamically evaluate the importance of features, even in the presence of features whose behavior changes over time. Experimental evaluations on datasets featuring whale and ship signals demonstrate the effectiveness and robustness of the proposed method.