cart_nodes_list suggested configurations?

[upadhyan]

What are the suggested configurations for cart_nodes_list? The original paper only mentions utilizing 8 configs, but doesn't mention what they are.

[KohlerHECTOR]

Hello, if you want to use the DPDTreeClassifier or Regressor you can use cart_nodes_list=(8,3,) . In that case, the decision tree algorithm will run reasonably fast. If you want to get something closer to an optimal decision tree you need to run something like cart_nodes_list=(100,) * max_depth . If you want to have the same results as the default sklearn decision tree you can run cart_nodes_list=(2,) .

If you want to use the AdaBoostDPDT classifier you should keep the single estimators parameters low like cart_nodes_list=(4,3,) .

Essentially, if cart_nodes_list=(i, j, k, l, ..., m) then it means DPDT will consider i candidates splits for each node at depth 0, j candidate splits for each node at depth 1, k candidates splits for each node at depth 2, and so on until depth m<= max_depth

[upadhyan]

Thanks! so to clarify, (8,3, ) means that DPDT will consider 8 splits at depth 0, 3 at depths 1 to max_depth?

[KohlerHECTOR]

Hello, to be extra precise (8,3,), it means that the candidate splits at depth 0 will be generated by calling the greedy algorithm from sklearn with max_leaf_nodes=8 (8-1 internal nodes in the greedy tree => 7 candidates), then at depth 1, the splits are generated by calling the greedy algorithm again with max_leaf_nodes=3 (3-1 internal nodes in the greedy tree => 2 candidates). So 7 candidates at depth 0 and 2 at depth 1 :) . Here is the exact part of the code of DPDT that generates the splits: https://github.com/KohlerHECTOR/DPDTreeEstimator/blob/d097719eb8d6299fc38ef901275cf8f3fdb0a598/dpdt/dpdt_classifier.py#L433C13-L452C18

[upadhyan]

Thanks for the help! This helps a lot.