I have been poking around the Rossmann Store Sales competition and reading code from the winners, as well as musing about the nature of deep learning and times series-style data in general. Please excuse a bit of meandering post!
Deep learning works well with non-rectangular datasets, such as sequences, text…and DL works well in computer vision over images…etc.
But when it comes to the classic “row/column, x number of (assumedly linearly separable) features, and a target” dataset…is there any point in trying to look for techniques past a multilayered perceptron? What deep learning models aside from a MLP are capable of exploiting the structure of this data?
Looking at the Rossman code that is still up on Kaggle, I see a lot of XGBoost…hummm…perhaps trying to apply DL to problems that can be address with more classic ML techniques is a quixotian task.
I’d love to hear anyone’s thoughts out there. Where can we go with rectangular data, past MLP?