UMAP + HDBSCAN grid search produces single dominant topic with few outliers

[powerhorse1986]

Hi all,

I have been experimenting with BERTopic on a corpus of 3,041 abstracts and noticed a puzzling pattern during hyperparameter tuning. I am hoping someone can shed some light on what is happening.

I performed a grid search over three parameters: n_neighbors and n_components (UMAP) and min_cluster_size (HDBSCAN). For certain parameter combinations, nearly all documents were assigned to topic 0 (e.g., 2,810 out of 3,041 abstracts), with very few outliers. Despite this collapse, the evaluation metrics (e.g., [insert metrics, e.g., coherence, silhouette score]) looked surprisingly good.

My questions are:

1. Why would most documents collapse into a single topic while the evaluation metrics still appear strong?
2. Is this a known failure mode of the UMAP + HDBSCAN pipeline under certain parameter combinations?
3. How should I interpret or guard against this pattern when using grid search to select hyperparameters?

Any hints or pointers would be greatly appreciated. Thank you!

Best regards,
Li