UMAP Parameters for Visualization

[eschaffn]

Hi there!

Two of the visualizations, visualize_topics() and visualize_documents() both use a 2d-reduced version of embeddings, visualize_topics() has a UMAP n_neighbors = 2, while visualize_documents() uses n_neighbors = 10.

My question is should these parameters be varied based on certain attributes of the fitted model/data? For example, should the n_neighbors of visualize_topics() be a function of the number of topics in the model, does this make mathematical sense? Should the same be applied to visualize_documents(), or should the parameter be scaled baed on some other property; not at all?

I've been doing my best to read the math behind UMAP and the n_neighbors parameter seems like it should vary a bit more.

Thanks

[MaartenGr]

My question is should these parameters be varied based on certain attributes of the fitted model/data? For example, should the n_neighbors of visualize_topics() be a function of the number of topics in the model, does this make mathematical sense? Should the same be applied to visualize_documents(), or should the parameter be scaled baed on some other property; not at all?

That's not easily done to make this a factor of the data itself since we know certain characteristics of the embeddings in low dimensional space only after reducing them in their dimensionality. Otherwise, there wouldn't have been a need for this parameter in the first place.

Ideally, it would be great if the parameters would be a function of the data but since the input data can vary wildly (embedding size, distribution of values, number of datapoints, etc.) there isn't a straightforward way to make sure all parameters are perfectly tuned towards the data.

[eschaffn]

I'd like to use these plots to get an idea of the metric I'm working to develop here: #2061

I'm setting the radius of the bubbles in visualize_topics() to the variance metric, and want to see if that correlates to an increase in spread in visualize_documents(). Would it be a good idea to keep the n_neighbors the same? And should it be lower (2?), or higher (10?)?

[MaartenGr]

That depends highly on the number of datapoints that you have. Since the number of topics are generally low (<200), I choose n_neighbors=2 since it would not run into any issues with very small number of topics.