Neural Network Ensembling with Random Features

This paper presents a novel approach to classification for tabular data. While neural networks excel at capturing nonlinear relationships within features, they often lack the robustness of ensemble methods. To address this limitation, we propose a model that combines the nonlinear learning capabilities of neural networks with the resilience of ensemble methods. Previous research on ensemble neural networks has shown promising results, although the lack of independence among models within the ensemble has been a challenge. Our approach aims to overcome this limitation by introducing a set of independently selected subsets of features to feed into a single neural network and ensembling the results. Experimental results demonstrate the effectiveness of this approach compared to baseline models on diverse datasets.