Should I perform customer segmentation before performing churn prediction?

Question

Imagine a company with multiple lines of revenues coming from diferent products, but all customer can access these different products through the same account and the same online platform. My goal is to predict the churn for each customer.

Should I perform customer segmentation into clusters and build a churn prediction model for each segment?

The straight foward path would be to get all relevant features for all customers and try to predict the churn for all of them. The different bahaviors and personae will theoretically reflect on the features.

What I do expect to get from the clustering are the personae this company intends to serve. In this case, a churn will be defined as a custumer who stops consuming any product or close its account. My objective here is to build a churn prediction model according to each pearsonae behavior pattern and I expect from this approach to have better features for each because now we understand them more and then more acurate models. This would also lead to more personalized business strategies to avoind the churn.
But I don’t know if this is a good practice and if I should try the straight foward approach or a different one.

Answer 1

There is no direct reason to segment your customers before estimating their churn probability, unless you know of business reasons for the segments to behave in a substantially different way with regard to churn, or the business users of the model for the different segments will have differing requirements (e.g. on model interpretability). Otherwise, there is nothing that a well-trained churn model could not do that two models (clustering and churn classification) would.

You can actually see for yourself if the approach brings any benefit - add the segmenting features to the churn model input and check whether they have important influence on its outcomes (they are high in the decision tree, have high feature importance, etc.) If they do, segmenting the customers and training separate churn models might bring benefits, especially if you have strict requirements on model interpretability.