Ensemble deep generalized eigen-value random vector functional link network for classification problems

Random vector functional link neural networks have been widely used across applications due to their universal approximation property. The standard random vector functional link neural network consists of a single hidden layer network, and hence, the generalization suffers due to poor representation of features. In this work, we propose ensemble deep generalized eigen value proximal random vector functional link (edGERVFL) network for classification problems. The proposed edGERVFL improves the architecture twofold: generating a better feature representation via deep framework, followed by the ensembling of the base learners, composed of multilayer architecture, to improve the generalization performance of the model. Unlike standard RVFL-based models, the weights are optimized by solving the generalized eigenvalue problem. To showcase the performance of the proposed edGERVFL model, experiments are conducted on diverse tabular UCI binary class datasets. The experimental findings, coupled with the statistical analysis, indicate that the edGERVFL model outperforms the provided baseline models.