We can try to optimize the tree's depth in order to maximize F1 or recall. In order to do so, we will experiment with depths in the range of [3, 11] on the train set.
The following graph depicts the F1 score and recall for the various maximum depths, both for the original and filtered datasets:
Here, we observe that for a maximum depth of 5, F1 and recall are optimized for the filtered dataset. Furthermore, recall is optimized for the original dataset as well. We will continue with a maximum depth of 5 as trying to further optimize the metrics can lead to overfitting, especially since the number of instances relevant to the metrics is extremely small. Furthermore, with a maximum depth of 5, there is an improvement both in F1, as well as in recall, when the filtered dataset is used.