1
$\begingroup$

guys!

I have a dataset with a bunch of costumer-behavior features and the output being "Churned"/"Not churned". I applied a simple Random Forest Classifier and got a nice performance.

With this, I can predict whether or not a given user will churn. But I need to understand what are the patterns among churned users and patterns among non-churned users. How could I achieve that? (Where I could present something like "Usually, users that churn do this, that and that")

PS: No need for a full explanation, I'd happy enough if you give me some directions to what to study so I can achieve this knowledge

Many thanks in advance!

Improve this question
$\endgroup$
1
  • $\begingroup$ You'd have to define what "patterns" mean to you to get a specific answer. $\endgroup$
    – Sean Owen
    Mar 25, 2016 at 18:08

1 Answer 1

Reset to default
1
$\begingroup$

A couple good options would be to look at a Feature/Variable importance plot for your RF model. Alternatively, depending on the model you could try extracting a couple individual trees from the model and examining them. However, these methods wouldn't be definitive; i.e. determining what variables are strong predictors for churn does not mean that they have a causal impact on churn, and an individual tree may be biased and not representative of the aggregation output presented by the RF model. To determine causation, you could use these methods as a starting point to design a test.

Improve this answer
$\endgroup$
1
  • $\begingroup$ That was extremely helpful, thanks. Could you elaborate a little bit more about this idea of creating a test? How could this determine causation? Thanks again, sir! $\endgroup$
    – Rodrigo Araújo
    Mar 28, 2016 at 20:03

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