Hi experts out there -

I have a user behavior log (e.g. # of logins, send post..etc) and trying to come up with a churn prediction model. A part of the request that I was asked was to find the value, or a moment that a user is about to churn. For example, when a user's post is under 5 in day 2 then a user will likely to churn (so the marketing can do something). The retention drops severely on the 3rd day, so I organized the data by behavior+timestamps(every 6hour) like below and ran the logistic regression

churned(0 or 1)|userid|event A count_at_6h)|...|event B_count at timeline at 72h 
  • event count here was cumulative by timeline

The issue that I faced is...

1) Performance is really low: AUC is under 0.7.
2) Is it possible spot "the moment" that the user will churn from this(e.g. users with less than 5 posts will churn) from any modeling exercise?

I will appreciate if anyone can advise how I can possibly proceed from here.. thank you in advance...

  • 1
    $\begingroup$ You either have to break down the event of churning into time cohorts or you need to use a specialised method such as Survival analysis. $\endgroup$ – Digio Mar 30 at 9:50
  • $\begingroup$ 2) so you can do discrete survival analysis with logistic regression. eg you split your time into 6 hour chunks and predict whether customer will churn in next 6 hours (assuming hasn't churned in preceding time periods). now you just have to decide on the inputs ( eg time, cumulative counts etc & interactions [ a standard set of variables for churn is RFM models..]). [ps you should look at the probability estimates rather than AUC - marketing should do cost benefit based on probabilities] $\endgroup$ – seanv507 Mar 30 at 14:22

Your Answer

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

Browse other questions tagged or ask your own question.