0
$\begingroup$

I'm currently working in a problem, where I think a supervised clustering approach might be a good candidate, but I'm not sure and haven't really worked with such scenario before. Let me break it down:

I'm working with a supervised scenario: I have some financial data and an associated probability of risk derived from another model. What I'd like to do is use that probability as a label and run a clustering algorithm to categorize the data according "partly" to the associated risk. That is, I want the algorithm to do a good job both at clustering the data on related features (proximity), but with the constraint that the associated risk is similar.

There might be better approaches to what I have in mind, which is why I'd be very happy to get feedback or other suggested methods, thanks.

$\endgroup$
4
$\begingroup$

it sounds like you're doing two things, but really you're just doing one:

you could think of the feature derived from this model as something which has to be compared to clustering results in some sort of ensemble fashion. I think a better approach would be to use the other models output as one of the inputs to the clustering algorithm. Then you can tune the other models weight on the clustering algorithm as a hyperparameter.

Check this out, you can set weights to k-means to make features more or less important. So if you just prepend the model output onto your current vector of inputs, then you can change how important the models output is vs. your other features with a weighting factor.

$\endgroup$
2
  • 1
    $\begingroup$ Thanks, this is actually were I-m heading. I'm realising there's not really a feasible approach to do supervised clustering in anyway having an actual probability as output. Unless I bin it, which could perhaps be an option. But I'm currently more convinced by this idea, adding it as a feature and running some unsupervised clustering algo. Thanks $\endgroup$
    – yatu
    Oct 6 at 19:32
  • $\begingroup$ Take into account that k-means should not be applied with all data types (e.g. categorical/not continuous data); take also into account that your model output should be calibrated to consider it a realiable predictor, and also how you use it to interpret predictions later on... $\endgroup$
    – German C M
    Oct 7 at 5:18
0
$\begingroup$

It sounds interesting, but maybe you could first try to plot the distribution of the features of interest for, on one hand, clients predicted with low risk, and on the other hand with high risk. It might maybe give you a first direct view of the most frequent values for each possible feature of interest (which could be the most important features reported by the classification model which gave the risk score).

On the other hand, you might be interested in looking at model explainability with SHAP values to interpret info of interest about why the model scored that way.

$\endgroup$
1
  • 1
    $\begingroup$ Hi German. This sounds like good advice. But neither using shap nor plotting distributions will get me an automated to assign a probability to new coming samples. I did use shap btw for further interpretability on the first model. Thanks $\endgroup$
    – yatu
    Oct 6 at 19:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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