Processing dates

Dates are common in databases and, especially when processing the forecasting of future estimates (such as in sales forecasting), they can prove indispensable. Neural networks cannot process dates as they are, since they are often expressed as strings. Hence, you have to transform them by separating their numerical elements, and once you have split a date into its components, you have just numbers that can easily be dealt with by any neural network. Certain time elements, however, are cyclical (days, months, hours, days of the week) and lower and higher numbers are actually contiguous. Consequently, you need to use sine and cosine functions, which will render such cyclical numbers in a format that can be both understood and correctly interpreted by a DNN.

Getting ready

Since we need to code a class operating using the fit/transform operations that are typical of scikit-learn, we import the BaseEstimator and TransformerMixin classes from scikit-learn to inherit...