Gradient Boosting from Concept to Observe (Half 2) | by Dr. Roi Yehoshua | Jul, 2023

Use the gradient boosting lessons in Scikit-Study to unravel completely different classification and regression issues

Within the first a part of this text, we offered the gradient boosting algorithm and confirmed its implementation in pseudocode.

On this a part of the article, we’ll discover the lessons in Scikit-Study that implement this algorithm, focus on their varied parameters, and exhibit the right way to use them to unravel a number of classification and regression issues.

Though the XGBoost library (which will probably be coated in a future article) supplies a extra optimized and extremely scalable implementation of gradient boosting, for small to medium-sized knowledge units it’s typically simpler to make use of the gradient boosting lessons in Scikit-Study, which have a less complicated interface and a considerably fewer variety of hyperparameters to tune.

Scikit-Study supplies the next lessons that implement the gradient-boosted choice timber (GBDT) mannequin:

1. GradientBoostingClassifier is used for classification issues.
2. GradientBoostingRegressor is used for regression issues.

Along with the usual parameters of choice timber, resembling criterion, max_depth (set by default to three) and min_samples_split, these lessons present the next parameters:

1. loss — the loss perform to be optimized. In GradientBoostingClassifier, this perform will be ‘log_loss’ (the default) or ‘exponential’ (which can make gradient boosting behave just like the AdaBoost algorithm). In GradientBoostingRegressor, this perform will be ‘squared_loss’ (the default), ‘absolute_loss’, ‘huber’, or ‘quantile’.
2. n_estimators — the variety of boosting iterations (defaults to 100).
3. learning_rate — an element that shrinks the contribution of every tree (defaults to 0.1).
4. subsample — the fraction of samples to make use of for coaching every tree (defaults to 1.0).
5. max_features — the variety of options to think about when looking for the most effective cut up in every node. The choices are to specify an integer for the…