Optimized Multi-Scale Convolutional Neural Network-Wavelet Feature Fusion With Reinforcement-Based Explainable Selection

Abstract

The escalating complexity of modern datasets demands sophisticated techniques for effective feature extraction, representation, and selection, as treating distinct data sources equally often hinders optimal performance. This study introduces a framework for Multi-Scale Feature Fusion and Adaptive Selection, transforming limited, fragmented information into meaningful insights, namely, AttenFusion-DualRL. The proposed strategy integrates a multi-stage pipeline: starting with data preprocessing for MNIST and Fashion-MNIST, followed by VGG16 feature extraction to leverage pre-trained visual knowledge. To enrich these with localized frequency information, we apply the Discrete Wavelet Transform (DWT). Subsequently, variance-based feature selection reduces dimensionality by retaining discriminative features. The core innovation lies in Cross-Domain Attention Fusion, which dynamically aligns and combines features from heterogeneous datasets. Finally, a Reinforcement Learning (RL) agent with a dual reward function optimizes feature selection by simultaneously maximizing classification accuracy, promoting sparsity, and enhancing interpretability through SHAP-guided rewards. Experiments on benchmark datasets demonstrate that the proposed strategy achieves higher classification accuracy with fewer features, showcasing improved model robustness. This approach promises significant future applications in image recognition, federated learning, and other domains that require efficient, explainable processing of diverse information sources.