1. Introduced HDBSCAN auto leaf/eom mode and auto min-sample selection
2. Fixed dependencies

Reduce with embeddings

Two topic reduce methods are introduced:

1. Reducing with clustering the topic vectors which created with document vector averaging. GMM is used for clustering and this reducing method works iff clustering method is GMM.
2. Reducing with HDBSCAN_flat. This method use hierarchical structure of HDBSCAN and search for an epsilon value to satisfy desired topic number to reduce.

What's Changed

• Embeddings based topic reduction + get rid of ST by @dopc and @zafercavdar in #13

Full Changelog: v1.0.5...v1.0.6

Keyword selection with new c-TF-IDF, custom embedding model and MMR

• c-TF-IDF update

• custom embedding model parameter, mmr update

• updated mmr_keywords() method for one topic usage

• updated embedding model as mmr_keywords() method parameter instead of bertopic class parameter

• Fixed calculate probabilities condition

New clustering method GM & block outlier merge

• clustering_method parameter introduced, GM method implemented

• topic index mismatch fixes, it was happening when clustering method does not produce outlier class

• -1 merge blocked, outlier class cannot merge and cannot be merged into

• topic number of reduce method is fixed for GMM clustering method

• reduce_topic topic number fix, topic number cannot be reduced 0 or 1 anymore

• Fixed HDBSCAN probs bug when it finds only -1 class

Outlier probability calculation and umap seed

• umap seed added
• outlier topic probability calculation added
• typo fixes in _append_outlier()