Consistency between initial clustering and approximate_distribution() similarity calculations

[Eric-Fithian]

When using approximate_distribution(), should the similarity calculation method match the approach used in initial clustering for consistency?

Specifically:

• Initial BERTopic clustering uses embeddings → clustering → c-TF-IDF topic representations
• approximate_distribution() can use either c-TF-IDF (through Bag-o-Words?) or embeddings (use_embedding_model parameter)

If documents were originally clustered based on semantic similarity in embedding space, could using c-TF-IDF without embeddings for distribution calculation lead to inconsistent results? Or does the method automatically handle this appropriately?

[MaartenGr]

When using approximate_distribution(), should the similarity calculation method match the approach used in initial clustering for consistency?

No, it is a different technique and appropriately named "approximate..." to indicate that it is not 1:1 with the initial clustering.

Initial BERTopic clustering uses embeddings → clustering → c-TF-IDF topic representations

Note that it is actually embeddings -> dimensionality reduction -> clustering -> Bag-of-Words -> c-TF-IDF

If documents were originally clustered based on semantic similarity in embedding space, could using c-TF-IDF without embeddings for distribution calculation lead to inconsistent results?

Yes and no. You are, to a certain extent, comparing apples to oranges. Specifically, the initial clustering of BERTopic is done on a document-level whereas the approximate_distribution then assigns the topics as a distribution within the documents. The two tasks are related but not exactly the same.

You indeed would want to expect that the most common topic in approximate_distribution matches that of the initial clustering but that will not always be the case considering this method is not primarily guided by the initial clusters.

[Eric-Fithian]

Thank you for the clarification. That makes sense. Does this mean that when the use_embedding_model flag is set that it is a 1:1 comparison for each token set to each topic cluster?

[MaartenGr]

It uses cosine simility to check between the embeddings of the token set to each topic embedding.