Table 5.

Comparisons for known, novel and all classes over three different iterations

Dataset	Known			Novel			All
	Ours	(38)	(36)	Ours	(38)	(36)	Ours	(38)	(36)
BM-CITE	0.82 ± 0.004	0.69 ± 0.015	0.64 ± 0.010	0.67 ± 0.046	0.74 ± 0.012	0.48 ± 0.009	0.79 ± 0.031	0.74 ± 0.018	0.54 ± 0.011
LUNG-CITE	0.89 ± 0.002	0.54 ± 0.010	0.42 ± 0.008	0.74 ± 0.004	0.73 ± 0.014	0.49 ± 0.006	0.87 ± 0.001	0.72 ± 0.016	0.46 ± 0.010
PBMC-Multiome	0.95 ± 0.004	0.78 ± 0.010	0.34 ± 0.009	0.76 ± 0.053	0.68 ± 0.012	0.69 ± 0.011	0.83 ± 0.028	0.72 ± 0.012	0.48 ± 0.010
PBMC-TEA	0.68 ± 0.008	0.73 ± 0.016	0.71 ± 0.010	0.72 ± 0.024	0.64 ± 0.013	0.32 ± 0.009	0.81 ± 0.015	0.73 ± 0.017	0.38 ± 0.010
PBMC-DOGMA	0.90 ± 0.015	0.72 ± 0.013	0.37 ± 0.010	0.50 ± 0.040	0.47 ± 0.014	0.44 ± 0.012	0.76 ± 0.038	0.63 ± 0.016	0.39 ± 0.010

For each dataset, 50% of the classes are regarded as known and the rest 50% as unknown. In addition, 70% of the known classes are labeled, with the unknown and the rest of the known all considered unlabeled. Best results are shown in bold.