In this new book release, at the top of page 51 the authors mention that to do deep learning on time series tabular data the developer should structure the tensors such that the channels represent the time periods.
For example, with a dataset of 17 features where each row represents an hour of a day: the tensor would have 3 dimensions,
x - the 17 features
y - the # of days
z - the 24 hours in each day
So each entry in the tensor would represent that day/hour.
Is this necessary to capture time series elements? Would the DNN not learn these representations simply by breaking up the date column into: day, hour?