How do you detect or control dangerous usage of variables and their transformation using featuretools?

Question

When you apply the transformation year, month, or day on a date of a purchase you could very easily run into a problem!

Imagine your purchase is in 2018 but you want to apply a model on data of 2019. the Model is developed using features automatically generated by featuretools including simple transformation like year.

The problem here is that the model does not know 2019!? Meaning the model is not general valid to new data.

score 0 · Answer 1 · answered Jul 01 '19 at 18:46

0

I do not believe the case you provided is a problem. The feature engineering calculation of extracting the year, month, or day of a date is allowed on the news year's data. As long as you don't training your model on future data you're trying to predict, you will be fine since you aren't leaking anything about the future into your training process.

answered Jul 01 '19 at 18:46

Max Kanter

2,006
6
16

Think of time series problems – Filip Floegel Jul 01 '19 at 22:14
Also consider time Windows for your features. When you have a time window of 4 years... the oldest year you do not have in the new data set with the time shift of one year – Filip Floegel Jul 01 '19 at 22:40

How do you detect or control dangerous usage of variables and their transformation using featuretools?

1 Answers1