This PR fixes a bugged edge case in DecisionTree.findBestSplit() where training would crash if when evaluating a new split all randomly chosen mtry columns contained only a single unique value. In that case, no split can be made without the right child containing no observations, and desired asserts would fail. The function has been extended to return -1 when this is case, which DecisionTee.fit() handles by not splitting that node further.
This PR fixes a bugged edge case in
DecisionTree.findBestSplit()where training would crash if when evaluating a new split all randomly chosenmtrycolumns contained only a single unique value. In that case, no split can be made without the right child containing no observations, and desired asserts would fail. The function has been extended to return-1when this is case, whichDecisionTee.fit()handles by not splitting that node further.