Explainability analysis based on attribution features for optimizing automatic modulation classification

With the explosive growth of Internet of Things devices, wireless communication systems face significant challenges in achieving wide-area coverage and service continuity. In highly dynamic and heterogeneous networks, modulation schemes exhibit pronounced time-variability and complexity, and traditional automatic modulation classification methods are severely limited under complex channel conditions and low signal-to-noise ratios. Deep learning can significantly improve recognition accuracy. However, its black-box nature and lack of interpretability hinder reliable model optimization and deployment in safety-critical scenarios. To address this, we propose an explainability-driven optimization framework that integrates feature attribution analysis and attention mechanisms to enhance model reliability and interpretability. Four mainstream interpretability methods (IG, DL, LIME, and SHAP) are applied to amplitude-phase and in-phase-quadrature feature domains, and the robustness and effectiveness of attribution features are evaluated via sliding window and feature deletion experiments, as well as misclassification case studies. Based on the selected effective features, a feature adjustment module and attention mechanism are introduced to guide the model’s focus on key features. An Explainability Metric for Modulation is further constructed to quantitatively assess the consistency between attribution results and physical signal characteristics via constellation-domain alignment analysis. Experimental results demonstrate that the framework improves interpretability and reliability without modifying the model architecture or introducing additional information, with recognition accuracy increasing by approximately 10 % for CNN and 6 % for LSTM, and the optimized LSTM achieving over 92 %, providing a practical and effective solution for automatic modulation classification in complex scenarios.