0
$\begingroup$

Basically, what I'm curious about, are there any methods in machine learning to make the model take into account events that happen in real time that affect the data points during that time period. Then, we can have the model put less weight on those data points or something like that in the fitted model.

For example, let's say we are making a customer churn prediction model for Netflix. We have training data for the last 1 year. We have the metadata and behavioural data of customers for every month as features, and use these to predict whether the customer will churn. That is typically the standard way of doing, but would it make sense that Covid would have affected customer behaviour and probably decrease churn of customers during the covid time. These customers could be very different to customers after Covid. In addition to that, there could be updates to Netflix that affected customers behaviour.

Would it be too complicated to teach the model to treat certain customers differently based on the date in the training of the model? I'm guessing it would and probably cause overfitting. It might be better just to generalise the model, but I'm curious what others think.

$\endgroup$

1 Answer 1

0
$\begingroup$

You could add more data points for time periods that are most recent, or remove data from older timestamps.

Using weights for data may not be a good idea: using a weight would imply increasing the learning rate for a time period, which often leads to divergence.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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