Parameter n in Train/Test split step

[djsmith17]

Hello there! I have a question regarding an analysis step seen at line 42 of generate_umap.py. At this step the code performs a train/test split of the data before performing the UMAP. I noticed two things about this function.

1. The value of n applied to the function is 2500. From my understanding of the train/test split this will find 2500 useable cells for the training sample and an additional 2500 useable cells for the testing sample. Does this value have significance for the UMAP calculation? Why was this number chosen over another? Are there decisions/pitfalls to consider when choosing this value of n?
2. I noticed you used a train/test split of 50/50. Could you share how you came to choose these split values? Is there something specific to your data that made you choose this split? Again are there decisions/pitfalls to consider when choosing the percentage split inherent to these sorts of experiments?

Thank you for your time and for sharing your thoughts

[alexbaras]

In these data there are so many cells to use - 2500 should be enough for training. I think the only real concern would be if there were not enough cells used in learning the UMAP manifold.