How to use LSTM to make prediction with both feature from the past and the current ones?

Question

Suppose I have a data frame with 2 columns, which are sales and promotions. I want to predict the next day sales based on the past sales and promotion info of 3 days, plus the promotion to be applied at the next day? How do I process and reshape the dataframe? I mean if only previous promotions need to be considered, then, after some shifting, the data frame could be reshaped into (sample size, 3, 2), but it becomes a problem if I also need to consider the promotion at the next day. It is a pretty common issue, does anyone has any thought about this?

Answer 1

Have a look at this question. There is a nice discussion about how to implement this. The idea in to create a separated Input to your model and concatenate it AFTER the recurrent layer(s). Also in the Keras documentation, there is an example on how to build such models with a few lines of code.