How can we get probabilities for all clusters in transform function?

[suprateek-19]

Currently we only get the predicted class through concept_model.transform() Can we get the predicted probabilities for each cluster or the top n clusters?

[MaartenGr]

That is currently not implemented. However, you can use the internal hdbscan model (concept_model.hdbscan_model) to extract the probabilities using its approximate_predict or hdbscan.membership_vector functions.

[shilpiag123]

I get following error while trying to access above which is a known issue too. Any other way to get probability distribution across concepts for images? AttributeError: 'HDBSCAN' object has no attribute 'approximate_predict'

[MaartenGr]

@shilpiag123 You should use the it as follows:

import hdbscan
probabilities = hdbscan.membership_vector(cluster_model, embeddings)

Having said that, you will have to access the cluster model and also pre-calculate the embeddings. Instead, I would advise using BERTopic v0.15 instead which how now support for topic modeling with images very similar to Concept.