Table 7.
Improvement δ in accuracy for different iterative optimization settings on the breast cancer dataset.
| Linear [%] | KNN [%] | Bayesian [%] | SVC [%] | NN [%] | |
|---|---|---|---|---|---|
| Base accuracy | 96.487 | 96.838 | 93.289 | 97.539 | 98.103 |
| Diagonal | 0.317 | –0.140 | 1.232 | 0.352 | –0.246 |
| Spherical | 0.422 | 0.000 | 1.443 | 0.176 | –1.230 |
| MLE | –0.177 | 0.279 | 1.477 | 0.246 | –0.281 |
| Iterative grid search (δ) | 0.175 | 0.245 | 1.371 | 0.211 | –0.598 |
| Shift (δ) | 0.175 | 0.069 | 1.229 | 0.105 | –0.773 |
| Shuffle (δ) | 0.175 | 0.245 | 1.336 | 0.211 | –0.598 |
| Finer (δ) | 0.316 | 0.245 | 1.371 | 0.211 | –0.598 |
| Combined 1 (δ) | 0.069 | 0.315 | 1.336 | 0.211 | –0.563 |
| Combined 2 (δ) | 0.239 | 0.140 | 1.442 | 0.211 | –0.669 |
Every classifier benefited from the proposed optimization apart from the NN. Finer improved the linear classifier the most, Combine 1 the KNN, Combined 2 the Bayesian. Therefore, the best hyperparameter setting was dependent on the classifier. Diagonal optimization was better than the proposed optimization for the linear classifier, SVC and NN. Spherical optimization was better for the linear and Bayesian classifier. MLE optimization was better for the Bayesian classifier, SVC and NN.