Experiment: Sampling for Class Imbalance

tldr; Not worth it

Experiment

1. Pink: Over-sample score 1 and 6, under-sample the rest.
2. Yellow: Over-sample only.

Outcome

Performance for scores 1 and 6 doesn't meaningfully improve, as evidenced by the metrics shown below. In fact it gets worse when doing any sampling. Moreover, performance of other scores also degrades. Training takes longer due to more samples.

QWK Scores

Confusion Matrix

F1 Score

Recall

Precision