0
$\begingroup$

I want to cluster my data and show which features were used to define the clusters to show the structure in my data.

To explain the use case: Imaging I have data from many products and I want show the variation and structure within my data. As Input features I have an BERT embedding (created with the help of the description of the product), and other categorical and numerical data, as the price, production country, ...

So far i have difficulties to find an suitable method, as most methods are not usable to cluster the data (unsupervised) and explain which features contributed to the cluster.

First, I was thinking if I could recreate the embedding with all features but this wouldn't help for the explanatory part. So do you have any advice how to approach to this problem?

$\endgroup$

1 Answer 1

0
$\begingroup$

I think you can approach your problem by dividing it into two stages:

  1. Clustering: use whatever method you deem appropriate for your data, e.g. k-means, k-medoids, HDBSCAN, etc.
  2. Explainable classification: train an explainable multiclass classifier on your data, using the clusters as labels. Alternatively, train one explainable classifier for each of the clusters (i.e. positive class is belonging to the cluster, negative class is not belonging), for instance, one logistic regression classifier per cluster. Then, interpret the predictor's influence in each classifier.
$\endgroup$

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.