Number of estimated parameters in a tree ?

[lcrmorin]

For comparaison and notably (using AIC/BIC information criterion) I was wondering if there is a natural equivalent of number of trainable parameters for trees. Intuitively I would consider one parameter at each split (the level of the split) + one parameter for each final leaf (the value to predict), summed over all trees.

Is my calculation correct ? Is this info directly accessible in the library ?

[jameslamb]

I'm not qualified to comment on how you might calculate measures like AIC, intended for parametric models, on non-parametric models like the tree-based ones produced by LightGBM. Maybe others will have a comment on that.

But I can directly answer the question "is [information about all the splits for all trees] available in LightGBM?". Yes, it is.

• Python: use Booster.trees_to_dataframe(), see Add a function to plot tree with a case #4784 (comment) for an example
• R: use lgb.model.dt.tree(), see [R-package] Output structure of predict() #5223 (comment) for an example

There are also other options to examine the structure of a fitted LightGBM model, like Booster.dump_model() (available in several of the LightGBM libraries) to dump the model format to JSON.

[lcrmorin]

I've found an authoritative answer to the technical question on the Stats SE that extend to the way to evaluate a model. The issue seems to be closable.

[jameslamb]

Thanks for coming back to share this and close this issue!

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.