mars.train reports inaccurate performance scores? about mars HOT 3 CLOSED

To report classification metrics on clustering/unsupervised tasks, the optimal assignment problem needs to be solved first. This is a combinatorial optimization problem which Hungarian algorithm solves in polynomial time. This is the most standard approach for evaluating performance of clustering algorithms (e.g. http://proceedings.mlr.press/v48/xieb16.pdf). We report 5 different evaluation metrics and evaluation is definitely not inaccurate.

from mars.

Thanks for your reply. But now I have another question: Hungarian algorithm does not allow us to label clusters with cell types for a completely unseen data (data that we do not have any cell type information available/ or have no information about portion of cell types in this dataset), right? Because Hungarian algorithm can only achieve best mapping/assignment between clustering labels if we have the ground truth (but ground truth is unavailable for a real world unlabeled data).

from mars.

Yes, it is used just for evaluation when you have ground truth annotations. If you do not have annotations, you should run MARS with evaluation_mode=False. In that case, I suggest you find differentially expressed genes for the cluster you obtain with MARS and check whether they make sense/agree with some marker genes.

from mars.