0
$\begingroup$

Take a classification task where the goal is to predict which customers are going to buy headache pills in a given week. One of the features that could be created is, for example, how many times the customer bought these pills in the past 60 days. However, the 60 days cutoff is a bit arbitrary and doesn't encode information efficiently. For example, we lose any information of what happened more than 60 days ago. Also, it considers someone who bought pills yesterday as exactly the same as someone who bought them 60 days ago.

I would like to know what are some usual ways to create features like that addressing these issues.

$\endgroup$
  • $\begingroup$ One could use multiple windows, some sort of decay or a full LSTM node of a neural network. $\endgroup$ – usεr11852 Jun 23 '18 at 22:38
1
$\begingroup$

You can try to add weights decreasing in time, exponentially or so.You won't lost any information, but 60 days ago will be less important than 1 day ago.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.