About parameter k in JSON file about higashi

[ruochiz]

Hi,

the $k$ is actually the neighbor_num parameter in the JSON file.

The first 20 epochs are just using cell associated features such as total read count per cell, batch id etc to regress out the potential bias, while the 120 epochs are running for the actual step 1.

You can change the number of epochs for sure. But one thing I notice is that although the auc/aupr values are not changing much, the embeddings can still change a lot. To check that, you might notice there is an "update rate" printed out per epoch indicating the min/max change of the embeddings (normalized by the l2 norm). If you see those numbers are really small, then the model actually converges. The test results showed that 80~90 epochs are sufficient for the datasets we studied. 120 is just to be safe.

If you have a follow up question, I would be happy to discuss through the email. [email protected].