Labeling function to get a good mix of positive/negative labels.

[coderaing]

I was wondering what will the "label times" look like for a problem statement like "Likelihood of making a purchase in next 7 days". The data at hand is the Customer data and all the previous transactions (Orders).

This guide (https://compose.alteryx.com/en/stable/examples/predict_next_purchase.html) is almost similar, but with a very big difference, in that it focuses on whether a customer will buy a "particular" product, and not any product. It is because of this difference that the guide gets both positive/negative labels.

If i use the same labeling function (with a slight modification to return true if purchase happened in that slice), i will get all positive labels, because our original data contains only data where the user did purchase something!

One solution i can think of is using drop_empty=False in the lm.search, so i get all slices even though no purchases happened in that slice. If this is the right approach, i have 2 related queries:

1. Will this dataset be more biased towards Negative labels? Since there might be long periods when no purchase happened for a customer.
2. Since we are dealing with 7 days window, the order_date might randomly fall in any one of the days of the slice. Is this ok?

The inference will happen every 2-3 days and the data will be handed over to marketing team.

Sorry if this seems like a very trivial problem, somehow i was not able to find enough resources online on how to handle this.