Table 5.

Level 3 classification results for sub-classes of all six main classes

Class	Size	False positives	Carry over false positives	Precision	New precision	Recall	Carry over false negatives	New recall	ROC area
1.1	81	4	10	95.06	84.61	95.1	15	80.20	0.96
1.2	81	5	10	93.98	83.87	96.3	10	85.71	0.96
1.3	81	7	10	91.76	82.11	96.3	7	88.63	0.99
1.4	81	2	10	97.44	86.36	93.8	6	87.35	0.96
1.5	34	3	5	91.18	79.48	91.2	14	64.58	0.89
1.10	43	0	7	100	86	100	1	97.72	0.98
1.16	35	0	5	100	86.48	91.4	4	82.05	0.98
2.1	38	0	6	100	86	97.37	10	77.08	0.96
2.2	45	0	7	100	86.5	100	0	100.00	0.97
2.3	23	0	4	100	85.2	100	17	57.50	0.96
2.4	44	1	7	97.73	84.3	97.72	6	86.00	0.99
2.5	43	4	7	91.49	79.6	100	6	87.75	0.98
2.6	36	3	6	91.89	79.1	94.44	13	69.38	0.95
2.7	36	1	6	96.77	81.1	83.33	3	76.92	0.94
2.8	37	4	6	89.47	77.3	91.89	8	75.55	0.89
3.1	35	1	8	96.97	78.05	91.43	8	74.41	0.97
3.2	45	2	10	95.74	78.95	100		100.00	0.95
3.3	34	1	8	96.97	78.05	94.18	11	71.11	0.94
3.4	46	2	10	95.56	78.18	93.48	3	87.75	0.93
3.5	42	0	10	100	79.17	90.48	4	82.60	0.97
3.6	44	3	10	93.62	77.2	100	2	95.65	0.97
3.7	20	3	4	86.96	74.07	100	6	76.92	0.92
3.8	13	1	2	92.31	80	92.3	5	66.66	0.90
3.11	18	1	4	94.44	77.27	94.45		94.44	0.97
4.1	76	4	11	94.87	83.15	97.37	12	84.09	0.99
4.2	78	4	11	95	83.51	97.44	15	81.72	0.98
4.3	90	4	13	95.7	83.97	98.89	1	97.80	0.97
4.4	41	1	6	97.37	84.09	90.24	5	80.43	0.97
4.6	42	2	6	94.87	82.22	88.1	7	75.51	0.96
4.99	43	2	6	95.24	83.33	93.02	3	86.95	0.97
5.1	83	2	10	97.59	87.1	97.6	8	89.01	0.98
5.2	39	1	4	97.5	88.63	100	9	81.25	0.99
5.3	78	1	8	98.73	89.65	100	16	82.97	1.00
5.4	40	2	5	94.87	84.1	92.5	9	75.51	0.97
5.5	40	0	5	100	88.63	97.5	9	79.59	0.98
6.1	207	2	18	99.04	91.19	100	1	99.51	1.00
6.2	73	1	6	98.61	91.02	97.26	19	77.17	0.98
6.3	72	3	6	95.65	88	91.67	13	77.64	0.95
6.4	35	5	3	86.49	80	91.43	1	88.88	0.94
Overall	2072	82	290	95.67	83.01	95.34	287	82.670	0.96

Carry over false positives and negatives from level two classifier experiments are taken into account while calculating precision and recall in level three. The distribution of the carry over false positives is factored by the number of test sequences in the respective sub-classes, in order to conserve the sequence distribution