Classification Results |
---|
 |
AUC
|
Sensitivity
|
Precision
|
NaN
|
F-score
|
NaN
|
HSMC - 15 species
|
HSMC |  | 0.887 ± 0.214 | 0.945 ± 0.059 | 0 | 0.895 ± 0.179 | 0 |
HSMC (11-fold CV) |  | 0.916 ± 0.130 | 0.956 ± 0.037 | 0 | 0.930 ± 0.083 | 0 |
PhyLearn - 15 species
|
PhyLearn - NJ |  | 0.992 ± 0.007 | 0.954 ± 0.041 | 0 | 0.924 ± 0.099 | 0 |
PhyLearn - UPGMA |  | 0.860 ± 0.211 | 0.931 ± 0.064 | 0 | 0.873 ± 0.153 | 0 |
PhyLearn - 74 species
|
PhyLearn - NJ |  | 0.741 ± 0.237 | 0.846 ± 0.181 | 1 | 0.768 ± 0.181 | 1 |
PhyLearn - UPGMA |  | 0.684 ± 0.256 | 0.860 ± 0.174 | 2 | 0.741 ± 0.180 | 2 |
Multi-class
|
15 species | 0.992 ± 0.010 | 0.902 ± 0.170 | 0.944 ± 0.054 | 0 | 0.911 ± 0.124 | 0 |
74 species | 0.982 ± 0.042 | 0.851 ± 0.189 | 0.901 ± 0.121 | 0 | 0.863 ± 0.145 | 0 |
- In this table, the three strategies are abbreviated as 'HSMC', 'PhyLearn' and 'Multi-class', respectively. The results of these three strategies are reported in the upper, middle and bottom part of the table, respectively. The results of hierarchical single-label multi-class classification are based on the FAME tree resulting from the divisive clustering experiment. Only the 15 species data set was considered and 3-fold and 11-fold stratified cross-validation (CV) was performed. In the case of phylogenetic learning, two 16S rRNA gene trees were used as template: neighbor-joining (NJ) and unweighted pair group method with arithmetic mean (UPGMA). For PhyLearn, both the 15 and the 74 species data set were considered and all PhyLearn experiments were performed using 3-fold stratified CV. Also the flat multi-class experiments were validated by this CV strategy. In the three strategies, classification performance was evaluated based on the pooled test set. Metrics reported are the area under the ROC curve (AUC), sensitivity, precision and F-score. Based on a multi-class confusion matrix, statistics were calculated in a one-versus-other setting with averaging of the corresponding statistic over the different classes. Standard deviations are also reported. NaN denotes the number of classes that have resulted in a value ∞ (only in case of precision and F-score).