Deep vs. Diverse Architectures for Classification Problems

Abstract

This study compares various superlearner and deep learning architectures (machine-learning-based and neural-network-based) for classification problems across several simulated and industrial datasets to assess performance and computational efficiency, as both methods have nice theoretical convergence properties. Superlearner formulations outperform other methods at small to moderate sample sizes (500-2500) on nonlinear and mixed linear/nonlinear predictor relationship datasets, while deep neural networks perform well on linear predictor relationship datasets of all sizes. This suggests faster convergence of the superlearner compared to deep neural network architectures on many messy classification problems for real-world data. Superlearners also yield interpretable models, allowing users to examine important signals in the data; in addition, they offer flexible formulation, where users can retain good performance with low-computational-cost base algorithms. K-nearest-neighbor (KNN) regression demonstrates improvements using the superlearner framework, as well; KNN superlearners consistently outperform deep architectures and KNN regression, suggesting that superlearners may be better able to capture local and global geometric features through utilizing a variety of algorithms to probe the data space.