Table 6 Impact of sampling strategy (for PCA). SMOTE was executed on the training-fold-only, post-reduction vectors (128-D) to respect the manifold geometry.

From: An empirical evaluation of dimensionality reduction and class balancing for medical text classification

Sampler

Macro-F\(_1\)

Training Time (s)

Memory (GB)

SMOTE (after PCA)

0.87 ± 0.02

42 ± 3

4.1

ADASYN

0.83 ± 0.03

55 ± 4

4.6

Under-sampling

0.79 ± 0.04

38 ± 2

3.2

SMOTE before PCA

0.83 ± 0.03

58 ± 4

4.5

Back to article page