Finding attributes that make up dense clusters of fraudulent transactions

Question

I have data about purchases customers made in my website.
Some users later decline the purchase, a scenario I'd like to avoid.

I have lots of data about the purchases made in my website, so I'd like to find clusters of users who share similar attributes, and are dense in "decliners" type of users. I have labelled data about those users (as we know who later declined the payment). The problem is, How do I cluster them in a meaningful manner?

The data is a high-dimensional one, as we collect various metrics about our users when they're browsing, but we can restrict the search to specific attributes.

I have tried some naive clustering algorithms variations, but those turned out problematic because the clusters did not make a lot of sense and we couldn't find actions (e.g conditions to feed our main model) to take on them.

Any suggestions or references to papers about the matter are highly appreciated!

Answer 1

Clustering is typically defined as finding groups of related things when there are not labels. In your case, you have labeled data (i.e., declined or not declined).

A more useful way to frame your problem is as feature importance. Which features of users are most important in being associated with the declined label?