Search code examples
pythonmachine-learningxgboostlightgbm

Is the importance_type 'split' of lightgbm the same as the importance_type 'weight` in xgboost?

Is the importance_type 'split' of lightgbm the same as the importance_type 'weight' in xgboost?

In other words, are the following the same?

booster.feature_importance(importance_type = 'split') # for lightgbm

and

get_fscore(importance_type='weight') # for xgboost
Solution

  • Despite the slightly different wording, they are the same indeed.

    From the LightGBM docs:

    If "split", result contains numbers of times the feature is used in a model.

    From the XGBoost docs:

    'weight’: the number of times a feature is used to split the data across all trees.

    No coincidence that these importance types are the default choices in the two packages respectively.