Wavelet-RNN: A randomized neural network with wavelet-transform-based feature extension

Abstract

The random vector functional link (RVFL) network is a leading shallow randomized neural network (RNN) known for its simple architecture and fast training. However, conventional RVFL networks have limited capacity to capture localized and multi-scale features, which restricts their effectiveness in modeling complex data patterns. Moreover, their dependence on random feature mappings often leads to suboptimal representations, especially in real-world datasets. To overcome these limitations, we introduce the Wavelet-transformed RVFL network (Wavelet-RNN), which integrates wavelet decomposition to enhance feature extraction. By leveraging wavelet transformation, our model effectively captures both spatial and frequency-domain information, improving generalization and robustness. This approach enables a more accurate representation of localized patterns and multi-scale structures, which conventional RVFL networks often overlook. To assess the effectiveness of the proposed Wavelet-RNN, we conduct extensive empirical studies on 20 binary and 20 multiclass real-world datasets from the UCI repository. Experimental results demonstrate that Wavelet-RNN consistently outperforms both standard RVFL and state-of-the-art RNN variants, showcasing superior robustness and effectiveness across diverse datasets. © 2025 Elsevier B.V., All rights reserved.