should we reduce the dimensionality of topic_model.topic_embeddings_ ?

[Batchounet]

Dear Creator of the amazing BERTopic

I want to perform cosine similarity of the topic_embeddings to a list of labels. I found it to perform better than zeroshot (and faster !) for my use case. However, the embeddings in topic_model.topic_embeddings_ are 384 dimensional vectors, ie their dimension is not reduced using hdbscan. To my understanding, the cosine similarity could suffer from the curse of dimensionality because of that. Actually, ploting the max cosine similarity to my list of labels might suggest that, making most topics cosine similar to my labels to 0. 55 :

Should I add a dimensionality reduction step ? would it be possible to use directly the reduced embeddings for the topic_model ?

Again thanks for your work

[MaartenGr]

To my understanding, the cosine similarity could suffer from the curse of dimensionality because of that.

Actually, that's not entirely the case. Sure the curse of dimensionality has some influence but that is generally much less compared to other distance measures, like euclidean distance. There's a reason why you see cosine similarity (alongside dot product) used in embedding-based computations and that's because these distance measures work so well.

For the "highest precision" I would advise not reducing the dimensionality of the embeddings when using plain cosine similarity.

However, the embeddings in topic_model.topic_embeddings_ are 384 dimensional vectors, ie their dimension is not reduced using hdbscan.

Note that it's UMAP reducing the embeddings, not HDBSCAN.

[Batchounet]

Thank you very much. Yes, I meant UMAP.