Matched label returns NaN in metric calculation

[Jieran-S]

SpaceHack2023/metric/jaccard/jaccard.py

Lines 60 to 63 in 2e81727

	if not args.matched_labels:
	contingency_table = pd.crosstab(domains, groundtruth)
	row_ind, col_ind = linear_sum_assignment(contingency_table, maximize=True)
	domains = domains.map(dict(zip(row_ind, col_ind)))

Some metrics(MCC, Jaccard) require matched labels, if the labels are not pre-matched, the script will implement a matching algorithm (above). But when no. of domain label > no. of ground truth label, the resulted domains object has many NaN, leading to downstream error.

Can any metric people look into it and propose a potential fix?

[shdam]

Could this be fixed with pd.crosstab(dropna = False)? https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.crosstab.html

Also, isn't the intention to prevent over-clustering with a resolution optimization or similar, which would prevent no. of domain label > no. of ground truth label to be true?

[Jieran-S]

Yea agree...The issue also arise from model who dont convert resolution to n_cluster. But in case we want to investigate robustness of the clustering methods in the future it would be good to have this option imo