valendin/rfm2lstm

Time-variant features

Opened this issue · 1 comments

Very nice work! Would it be possible to include time-variant covariates that are conditional on a transaction? For example the product category as a dummy-encoded variable? In that case, how would the model deal with the 0s or missing values due to a lack of transactions?

Thanks Tim -- this would be easily possible, you just need to include the dummy-encoded variable as an additional input at each time-step. A missing value would just be another category in this case. The issue is during the holdout period, where presumably you don't have this information available -- what you can do is to have the model learn to predict the category as well with an additional output layer.