DARTOptions
Options & Configuration · AiDotNet.Models.Options
Configuration options for DART (Dropouts meet Multiple Additive Regression Trees).
For Beginners
Regular gradient boosting can overfit by making each new tree perfectly complement all previous trees. DART introduces randomness by “dropping out” some trees during training:
- Imagine you have a team of experts (trees) making predictions together
- In regular boosting, each new expert learns to fix exactly what the whole team got wrong
- In DART, when training a new expert, some existing experts are temporarily removed
- This forces the new expert to be more versatile, not just filling a specific gap
- The result is a more robust ensemble that generalizes better to new data
DART is especially useful when:
- You’re overfitting with regular gradient boosting
- You want more robust predictions
- You have enough computational budget (DART is slower than regular boosting)
How It Works
DART applies the dropout concept from neural networks to gradient boosting. During each boosting iteration, a random subset of existing trees is dropped, and the new tree is fit to the residuals considering only the non-dropped trees. This helps prevent overfitting and improves generalization.
Reference: Rashmi, K.V. & Gilad-Bachrach, R. (2015). “DART: Dropouts meet Multiple Additive Regression Trees”. AISTATS 2015.
Constructors
| Constructor | Summary |
|---|---|
DARTOptions() |
Properties
| Property | Summary |
|---|---|
DARTDropoutMode DropoutMode { get; set; } |
Gets or sets the dropout mode. |
DARTNormalization Normalization { get; set; } |
Gets or sets the normalization strategy after dropout. |
bool OneDrop { get; set; } |
Gets or sets whether to use one-drop strategy. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate (probability of dropping a tree). |
double FeatureFraction { get; set; } |
Gets or sets the fraction of features to consider for each split. |
double L2Regularization { get; set; } |
Gets or sets the L2 regularization strength for leaf weights. |
double LearningRate { get; set; } |
Gets or sets the learning rate (shrinkage factor). |
double MinSplitGain { get; set; } |
Gets or sets the minimum loss reduction required to make a split. |
double SampleDroppedTreesProbability { get; set; } |
Gets or sets the probability of adding a dropped tree back to the ensemble. |
double SubsampleFraction { get; set; } |
Gets or sets the fraction of samples to use for each tree. |
int MaxDepth { get; set; } |
Gets or sets the maximum depth of each tree. |
int MinSamplesLeaf { get; set; } |
Gets or sets the minimum number of samples required to split a node. |
int NumberOfIterations { get; set; } |
Gets or sets the number of boosting iterations (trees to build). |
int SkipDropoutIterations { get; set; } |
Gets or sets whether to skip dropout during the first k iterations. |
int? Seed { get; set; } |
Gets or sets the random seed for reproducibility. |