How to choose best parameters for ensuring local coherent structures with maximum spread

[PranayMehta]

I am working on survey text data and I want to implement clustering to group survey responses into multiple topics/clusters. For this, I use BERT model and transform each record into 768D array. Since I want to apply dimensionality reduction , I am trying out UMAP to see if 2D plan retains the structure.

Sample data points would look like

sentences = ['I am worried about the weather tomorrow' , ' I have a lot of homework', 'I am yet to take a covid19 vaccine']

In general, I have found that this library and algorithm is super helpful in helping me identify regions in 2d plane. As an example, there are a bunch of responses that talk about topic A and others about topic B and C and so on. My task is to keep these topics as isolated as possible.

However, I am a little confused with the number of parameters that are available in the UMAP method. @lmcinnes It would be super helpful if I can tune the UMAP algorithm to do so. As of now, I tried with several values of parameters for

1. neighbors
2. min_dist
3. spread

What I do not understand at the moment is

1. how does spread parameter work ? What is the range that this param can take?
2. What are the ranges of other n_neighbors and min_dist? Is the distance calculated according to the metric value passed?

Here are some of the plots that I got by varying the parameters

There are clusters that are forming well but then there is 1 big cluster that kind of gets clumped in the middle . any help is appreciated :-)