Skip to main content

         This documentation site is for previous versions. Visit our new documentation site for current releases.      

Configuring advanced settings for an adaptive model based on gradient boosting

Updated on July 5, 2022

Configure advanced settings that control the operation and learning of an adaptive boosting model. Properly configured adaptive boosting models are more likely to return relevant results.

Before you begin: Configure basic settings for an adaptive model. For more information, see Configuring settings for adaptive models. An adaptive model rule that is based on gradient boosting has a different set of parameters than a Bayesian adaptive model. The parameters for an adaptive model rule can be found under the advanced settings of the adaptive model rule. The parameters should only be changed with extreme caution, as the default settings are chosen to make the modeling technique both robust and powerful, so any change from these defaults could affect both the predictive power of the algorithm, and slow down the scoring time of the model.Note: The online nature of the adaptive model makes it impossible to do a hyperparameter optimization.
Advanced settings for an adaptive model based on gradient boosting
Settings tab showing the parameters and their default values.
  1. In the Learning rate field, enter an ETA (shrinkage factor) numeric value between (but larger than) 0 and 1.
    This parameter prevents over-fitting by slowing down the learning process of the model. A smaller value (closer to 0) means that each tree has a small weight on the prediction score of the model. This implies that more trees are necessary to get a better prediction performance. A higher ETA value (closer to 1) means that less trees are going to be learned, because the weight of individual trees is high. The recommended and default value is 0.3.
  2. In the Lambda field, enter a numeric value for regularization of the model on leaf weights. This value adds a penalty when the model complexity increases.
    Lambda is also used for the calculation of gain, which is the quality metric at the node level. Increasing Lambda will make the model split more conservatively, because the gain will be lower than the gain calculated without Lambda. The recommended and default value is 1.
  3. In the Gamma field, enter a numeric value for the complexity threshold that controls splits and prune.

    Gamma is the threshold for the minimum gain a node needs before we allow it to split. It is also a threshold for pruning, and defines the minimum gain a node needs to have to not be considered for pruning. A low gamma value will force the model to split or prune more often.

    The recommended and default value is 0.
  4. In the Minimum child weight field, enter a numeric value to constrain a possible node split.
    Minimum child weight designates the minimum sum of instance weight (hessian) needed in a child. If the tree partition step results in a leaf node with the sum of instance weight less than min_child_weight, then the building process will stop further partitioning. In linear regression, this corresponds to the minimum number of instances needed in each node. The larger the value, the more conservative the algorithm will be, forcing to not split on that node. The recommended and default value is 1.
  5. In the Maximum trees field, enter the maximum number of trees that can be added to the model, between 1 and 500.
    For a model containing shallow trees (maximum tree depth), maximum trees should allow the creation of a high number of trees, so that a new tree is created whenever possible. This will also enable generalization of the model. Increasing the value will increase the scoring time of the model. The recommended and default value is 50.
  6. In the Maximum tree depth field, enter the maximum depth of a single tree, between 1 and 14.
    The higher the value, the more complex the tree. A low depth value results in shallow trees, because the trees will not grow more than the predefined value. This is desirable in the adaptive boosting model. The maximum tree depth parameter should correspond to the maximum trees value. Increasing the value will increase the scoring time of the model. The recommended and default value is 9.
  • Adaptive gradient boosting overview

    Self-learning predictive models in Adaptive Decision Manager play a crucial part in a Next-Best-Action strategy. Adaptive models predict the propensities of all the available actions and so provide highly personalized and relevant actions to each individual customer, achieving true 1:1 customer engagement.

Have a question? Get answers now.

Visit the Support Center to ask questions, engage in discussions, share ideas, and help others.

Did you find this content helpful?

Want to help us improve this content?

We'd prefer it if you saw us at our best. is not optimized for Internet Explorer. For the optimal experience, please use:

Close Deprecation Notice
Contact us