Cyclostationary Feature Selection
Abstract
This article describes an effective procedure for identifying cyclostationary features that strongly influence the performance of certain standard machine learning prediction models. Cyclostationary features such as moments and cumulants extracted from raw electromagnetic signals are often utilized in detecting and classifying digitally modulated signals. However, not all features contribute equally to the outcome. In this work, we explore and implement SHAP (SHapley Additive exPlanations), a game theoretic method to determine the significance of each cyclostationary feature in predicting the modulation types associated with received signals. Using properly reduced feature sets obtained from SHAP, we demonstrate that models such as XGBoost and Random Forest could achieve classification accuracy comparable to the baseline with full feature sets. Furthermore, our empirical results indicate that a balanced choice of significant features could improve computational efficiency without compromising prediction performance. Using cyclostationary feature selection as a use case, we show that the suggested approach could be applied to a broader range of datasets and machine learning techniques to identify and quantitatively explain the factors that most likely influence prediction model results.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 22, 2023
- Accession Number
- AD1215967
Entities
People
- Anthony Tai
- Savannah Farney
Organizations
- Naval Surface Warfare Center Crane Division