1
$\begingroup$

I have some labeled data (0=didn’t cancel, 1=canceled) that I am creating a model for in my marketing class.

On top of predicting who is likely to cancel, I’d like to explore the possibility of trying different proactive retention strategies. I was thinking of running k-means on the training data where the label=1 and get, say, 4 clusters.

Is this the right way to go about this? I would basically end up with two models and run each customer through the binary classifier, and if it’s predicted to cancel, run the customer through the clustering model.

I’m not sure of this approach because k-means is an unsupervised learning method and I’m sort of helping it by feeding it just the customers in the positive class.

Please share your thoughts on this approach and any suggestions.

Improve this question
$\endgroup$
1
  • 2
    $\begingroup$ You may be way too optimistic on k-means ability to process your data in a meaningful way, and of the chances of finding a behavioral pattern, and not finding trivial customers groups (say, old males, old females, young males, young females). $\endgroup$
    – Has QUIT--Anony-Mousse
    Commented Aug 14, 2019 at 5:57

1 Answer 1

Reset to default
1
$\begingroup$

I'm assuming you've also observed some covariates for each customer beyond whether or not they canceled. Your suggested approach, then, would simply give you 4 groups of customers who canceled that are similar in terms of those other covariates. You could then examine the characteristics of those groups to see if they merit different retention strategies based on marketing theory. However, this would be a somewhat ad-hoc approach.

A more statistical way to approach the question would be this. First, can I tell which covariates determine whether or not a customer is likely to cancel? A logistic regression model could do this, for example. If you then suspect that customers who differ on a certain covariate $X$ are likely to cancel for different reasons, then you could include an interaction between $X$ and other covariates in the logistic regression model.

For example, suppose you're selling newspapers and you observe $Y =$ cancellation, $X =$ political affiliation (Democrat/Republican/Independent), and a measure of "liberalness" of the news customers read on a 1-10 scale $Z$. Then you might run the regression $Y \sim X + Z$ and conclude that Democrats are most likely to cancel but that liberalness doesn't really affect cancellation much. But then you might run the regression $Y \sim X + Z + XZ$, which includes an interaction term, and realize that higher liberalness decreases cancellation probability for Democrats, doesn't change it for Independents, and increases it for Republicans.

You might then conclude that you need to target your paper's articles more aggressively to the politics of its readers, which would be good for your circulation but probably bad for society.

Improve this answer
$\endgroup$
2
  • $\begingroup$ That’s a great approach. When you say my approach would be somewhat ad-hoc, do you mean because the strategy is being determined based on the clustering algorithm whereas in your approach, you’re using the coefficients to help develop the strategies? To be clear, the clustering model would score new customers based on what it learned during training, so that would be automated. $\endgroup$
    – Insu Q
    Commented Aug 14, 2019 at 1:00
  • $\begingroup$ Right--the issue with your clustering approach is that it does nothing to connect the attributes of customers to whether or not they cancel, so you have to do that part yourself. The regression methods I suggest here make that connection in an automatic and interpretable fashion. $\endgroup$
    – Sheridan Grant
    Commented Aug 14, 2019 at 21:49

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.