Visualize situations where the parameters of dimensionality reduction are inconsistent with those of the passed model

[superseanyoung]

When I use UAMP to reduce dimensions, I set n_neighbors=30,n_components=15, and then pass in BERTopic() with this UMAP. However, I want to use the visualize_document_datamap () method in the subsequent normalization, and then I introduce the reduced_embeddings, which I obtained through UMAP ().fit_transform. But the parameters I set there are n_neighbors=30,n_components=2, which is different from the parameters I passed into the BERTopic model, so what's the impact of that?

[MaartenGr]

There's always some mismatch between the actual procedure of fitting the model and the 2D representation since we cannot easily visualize in higher dimensionality than that. 2D visualizations are always an approximation of reality. That said, are you referring to something specific when you mention "impact"?

[superseanyoung]

My front and back parameters are different. Is there a logical error in this operation?Or maybe I can have the input model's parameters different from the parameters of my dimensionality reduction?

[MaartenGr]

My front and back parameters are different. Is there a logical error in this operation?

I don't think so. As mentioned a 2D representation is still a proxy of the original, so take that for what it is.

Or maybe I can have the input model's parameters different from the parameters of my dimensionality reduction?

The 2D visualization is an approximation of the underlying clustering process. As such, you will never get the exact same thing. That is something to take into account, but you for example can't just use 15 components for both the fitting and the visualization as 15 dimensions cannot be visualized that way.

Therefore, you take 2D representations but the underlying principle is the same, albeit with information loss. This is the suggested approach and although it is not an exact match (and unlikely ever will), I would not be worried about the difference in parameters here other than what I mentioned above.