Improve `support` metric for ratings other than 5

[LeoBorai]

Currently support metric for rating of 5 stars is far beyond counterparts.
This might be due to a lack of samples on other classes.

Find more examples for rating of [4;1] to improve this metric, perhaps by having
sets of 500 samples per class could help improve support.

              precision    recall  f1-score   support

           1       0.25      0.25      0.25         4
           2       0.50      0.22      0.31         9
           3       0.80      0.57      0.67        28
           4       0.42      0.30      0.35        50
           5       0.73      0.86      0.79       144

    accuracy                           0.67       235
   macro avg       0.54      0.44      0.47       235
weighted avg       0.65      0.67      0.65       235