Table 5.
The results of discriminant analysis variants and classification tree method when different ECOC coding designs have been used. Different coding designs are abbreviated as follows: one-vs-all (OVA), one-vs-one (OVO), ordinal (ORD), and ternary complete (TER). Quadratic discriminant analysis could not be evaluated due to nonpositive definiteness of covariance matrix. True positive rates can be found from the parenthesis next to true positive result and accuracy from the last column of the table.
| Method/class | Bad | Good | Semigood | ACC |
|---|---|---|---|---|
| OVA-LDA | 17 (41.5%) | 34 (45.9%) | 16 (27.6%) | 38.7% |
| OVO-LDA | 22 (53.7%) | 27 (36.5%) | 20 (34.5%) | 39.9% |
| ORD-LDA | 17 (41.5%) | 24 (32.4%) | 18 (31.0%) | 34.1% |
| TER-LDA | 16 (39.0%) | 32 (43.2%) | 20 (34.5%) | 39.3% |
| OVA-diagLinear | 19 (46.3%) | 59 (79.7%) | 11 (19.0%) | 51.4% |
| OVO-diagLinear | 16 (39.0%) | 58 (78.4%) | 16 (27.6%) | 52.0% |
| ORD-diagLinear | 19 (46.3%) | 39 (52.7%) | 15 (25.9%) | 42.2% |
| TER-diagLinear | 17 (41.5%) | 58 (78.4%) | 15 (25.9%) | 52.0% |
| OVA-pseudoQuadratic | 41 (100.0%) | 0 (0.0%) | 0 (0.0%) | 23.7% |
| OVO-pseudoQuadratic | 9 (22.0%) | 35 (47.3%) | 28 (48.3%) | 41.6% |
| ORD-pseudoQuadratic | 41 (100.0%) | 0 (0.0%) | 0 (0.0%) | 23.7% |
| TER-pseudoQuadratic | 9 (22.0%) | 31 (41.9%) | 31 (53.4%) | 41.0% |
| OVA-classification tree | 17 (41.5%) | 39 (52.7%) | 15 (25.9%) | 41.0% |
| OVO-classification tree | 19 (46.3%) | 50 (67.6%) | 30 (51.7%) | 57.2% |
| ORD-classification tree | 13 (31.7%) | 48 (64.9%) | 17 (29.3%) | 45.1% |
| TER-classification tree | 16 (39.0%) | 48 (64.9%) | 23 (39.7%) | 50.3% |