"Plain" train mode still build the oblivious tree about catboost HOT 3 CLOSED

According to the paper, it said the plain mode will build the tree by following the GBDT strategy. However, when I plot the tree trained with Plain mode, the structure of tree is oblivious tree not the decision tree.

Oblivious trees are a subtype of decision trees so no contradiction here. If you want CatBoost to build non-oblivious trees you can specify grow_policy training parameter as Depthwise or Lossguide.

Another confusing point is why the priornum used in each layer is different？

Because there can be several numerical features derived from categorical ones with different priors.

You can see the default sets of priors for each Ctr type here.

It is also possible to specify your own priors using training parameters.

from catboost.

According to the paper, it said the plain mode will build the tree by following the GBDT strategy. However, when I plot the tree trained with Plain mode, the structure of tree is oblivious tree not the decision tree.

Oblivious trees are a subtype of decision trees so no contradiction here. If you want CatBoost to build non-oblivious trees you can specify grow_policy training parameter as Depthwise or Lossguide.

Another confusing point is why the priornum used in each layer is different？

Because there can be several numerical features derived from categorical ones with different priors.

You can see the default sets of priors for each Ctr type here.

It is also possible to specify your own priors using training parameters.

@andrey-khropov Appreciate for your help. But one question you missed is how the leaf value calculated? In the tree_plot it shows as 0.071 but the prediction return the value 0.5166, and I am confused about how the leaf node value is converted into the predicted value？

from catboost.

how the leaf value calculated? In the tree_plot it shows as 0.071 but the prediction return the value 0.5166, and I am confused about how the leaf node value is converted into the predicted value？

1. You display only the single tree (specified by tree_idx), but the model typically contains many trees (up to the number of iterations but may have less than the specified maximum value of iterations if the overfitting detector and/or use_best_model are used). The number of trees in the model can be obtained from the tree_count_ attribute of the trained model.

2. The return value of predict will be the function of the sum of values of the selected leaves in all trees. Post-transformation depends on the specified prediction_type parameter.

from catboost.