ScaleGBMRegressor¶
- class getml.predictors.ScaleGBMRegressor(colsample_bylevel: float = 1.0, colsample_bytree: float = 1.0, early_stopping_rounds: int = 10, gamma: float = 0.0, goss_a: float = 1.0, goss_b: float = 0.0, learning_rate: float = 0.1, max_depth: int = 3, min_child_weights: float = 1.0, n_estimators: int = 100, n_jobs: int = 1, objective: str = 'reg:squarederror', reg_lambda: float = 1.0, seed: int = 5843)[source]¶
- Standard gradient boosting regressor that fully supports memory mapping
and can be used for datasets that do not fit into memory.
Gradient tree boosting trains an ensemble of decision trees by training each tree to predict the prediction error of all previous trees in the ensemble:
\[\min_{\nabla f_{t,i}} \sum_i L(f_{t-1,i} + \nabla f_{t,i}; y_i),\]where \(\nabla f_{t,i}\) is the prediction generated by the newest decision tree for sample \(i\) and \(f_{t-1,i}\) is the prediction generated by all previous trees, \(L(...)\) is the loss function used and \(y_i\) is the target we are trying to predict.
The regressor implements this general approach by adding two specific components:
The loss function \(L(...)\) is approximated using a Taylor series.
The leaves of the decision tree \(\nabla f_{t,i}\) contain weights that can be regularized.
These weights are calculated as follows:
\[w_l = -\frac{\sum_{i \in l} g_i}{ \sum_{i \in l} h_i + \lambda},\]where \(g_i\) and \(h_i\) are the first and second order derivative of \(L(...)\) w.r.t. \(f_{t-1,i}\), \(w_l\) denotes the weight on leaf \(l\) and \(i \in l\) denotes all samples on that leaf.
\(\lambda\) is the regularization parameter reg_lambda. This hyperparameter can be set by the users or the hyperparameter optimization algorithm to avoid overfitting.
- Args:
- colsample_bylevel (float, optional):
Subsample ratio for the columns used, for each level inside a tree.
Note that ScaleGBM grows its trees level-by-level, not node-by-node. At each level, a subselection of the features will be randomly picked and the best feature for each split will be chosen. This hyperparameter determines the share of features randomly picked at each level. When set to 1, then now such sampling takes place.
Decreasing this hyperparameter reduces the likelihood of overfitting.
Range: (0, 1]
- colsample_bytree (float, optional):
Subsample ratio for the columns used, for each tree. This means that for each tree, a subselection of the features will be randomly chosen. This hyperparameter determines the share of features randomly picked for each tree.
Decreasing this hyperparameter reduces the likelihood of overfitting.
Range: (0, 1]
- early_stopping_rounds (int, optional):
The number of early_stopping_rounds for which we see no improvement on the validation set until we stop the training process.
Range: (0, \(\infty\)]
- gamma (float, optional):
Minimum loss reduction required for any update to the tree. This means that every potential update will first be evaluated for its improvement to the loss function. If the improvement exceeds gamma, the update will be accepted.
Increasing this hyperparameter reduces the likelihood of overfitting.
Range: [0, \(\infty\)]
- goss_a (float, optional):
Share of the samples with the largest residuals taken for each tree.
If goss_a is set to 1, then gradients one-sided sampling is effectively turned off.
Range: [0, 1]
- goss_b (float, optional):
Share of the samples that are not in the goss_a percentile of largest residuals randomly sampled for each tree.
The sum of goss_a and goss_b cannot exceed 1.
Range: [0, 1]
- learning_rate (float, optional):
Learning rate for the gradient boosting algorithm. When a new tree \(\nabla f_{t,i}\) is trained, it will be added to the existing trees \(f_{t-1,i}\). Before doing so, it will be multiplied by the learning_rate.
Decreasing this hyperparameter reduces the likelihood of overfitting.
Range: [0, 1]
- max_depth (int, optional):
Maximum allowed depth of the trees.
Decreasing this hyperparameter reduces the likelihood of overfitting.
Range: [0, \(\infty\)]
- min_child_weights (float, optional):
Minimum sum of weights needed in each child node for a split. The idea here is that any leaf should have a minimum number of samples in order to avoid overfitting. This very common form of regularizing decision trees is slightly modified to refer to weights instead of number of samples, but the basic idea is the same.
Increasing this hyperparameter reduces the likelihood of overfitting.
Range: [0, \(\infty\)]
- n_estimators (int, optional):
Number of estimators (trees).
Decreasing this hyperparameter reduces the likelihood of overfitting.
Range: [10, \(\infty\)]
- n_jobs (int, optional):
Number of parallel threads. When set to zero, then the optimal number of threads will be inferred automatically.
Range: [0, \(\infty\)]
- reg_lambda (float, optional):
L2 regularization on the weights. Please refer to the introductory remarks to understand how this hyperparameter influences your weights.
Increasing this hyperparameter reduces the likelihood of overfitting.
Range: [0, \(\infty\)]
- seed (int, optional):
Seed used for the random sampling and other random factors.
Range: [0, \(\infty\)]
Methods
validate([params])Checks both the types and the values of all instance variables and raises an exception if something is off.
Attributes