The neural networks for structured datasets
So, considering research in this regard. As redes neurais tem um ótimo desempenho em dados não-estruturados, mas dados estruturados, modelos tradicionaias ainda são competitivos. The SFR-Net proposition was a study where the inductive bias consisted in provide an neural network expert for each feature and there is a block that learns the best set considering the features for each instance.
Why the neural networks are not so great for structured datasets considering that the architecture is so structured? Sure, there are some thought reasons, like the heterogeneity regarding the features, each one with a semantic meaning. For that, one neural network expert for each one, distinct representations, but just one signal in this regard.
Great challenge and interesting task in this regard.
Really great learning.