Wrong Medicine

A tabular classifier reports 61% training accuracy and 60% validation accuracy. A much smaller baseline on the same data gets 59% and 59%. Following scikit-learn's reading of a validation curve, both scores being low means the estimator is underfitting.

Select every change that attacks THIS problem. (Select all that apply)

Stop training earlier, since early stopping protects the validation score.

Raise the dropout rate, since dropout is the standard fix for a disappointing validation score.

Add features or interactions the current model cannot express from the columns it has.

Add capacity: deeper trees, more boosting rounds, or a larger hidden layer.

Weaken the regularization: lower the L2 penalty, raise the depth cap, lower min_samples_leaf.

Collect ten times more training rows, since more data always improves accuracy.