Redes neurais para dados tabulares: uma comparação empírica

Abstract

Deep Learning has undergone significant transformations in recent decades, making it the dominant strategy in modeling unstructured data, such as videos, audios, language, and images. Although it came to structured data, it was not able to overcome the popularity of more established methods such as Generalized Linear Models, Gradient Boosted Trees and Bagging. In recent years, promising models have emerged for tabular data that aim to adapt, for neural networks, innovations applied to unstructured data, such as the architecture of transformers, or that tried to simulate processes based on decision trees. This study is an empirical comparison of these models in databases of very different dimensions and sample sizes. It is illustrated by various areas, such as materials science, marketing, biology, and astronomy.