I have a model which calculates churn probabilities for e-commerce customers based on their historical activity data (no. of sessions, page views, purchases etc) using supervised learning with an AUC of 88%.

The task I have at hand is to find causes for churn. The analysis that I did has a 5 month observation period data where I look at causal metrics (service delivery, mode of service, returns made and the reasons for it, customer-support contact frequency etc) and try to predict customers who would churn (used a threshold on churn probability to convert to a binary variable). The model is a weak one with an AUC of 62% (using random forest) and there is no single variable which comes out to be a strong indicator for churn in this model. I am doing some feature engineering at present but my initial attempts haven't resulted in anything promising.

Is there a better way to do such an analysis?

  • $\begingroup$ How many features do you approximately have to choose your predictors from? Are you using all of them at the moment? $\endgroup$
    – Alex VII
    Commented May 18, 2018 at 12:35


Your Answer

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