fit_parameters

Description

This configuration is only available with the XGBoost algorithm.

When using XGBoost as a model, Verteego offers a way to combine the benefits of a global model with the benefits of a local model.

It provides a method to tailor the fitting process, combining the advantages of a global model with those of a local model. These parameters might include settings for handling imbalanced data, specifying the number of boosting rounds, or other XGBoost-specific options that influence how the model learns from the data.

To do so, an advanced method has been developed to pick the best number of estimators when predicting, for a given resolution level.

Options

The parameter resolution indicates the resolution used to optimize the number of estimators. gain_threshold specifies the minimal gain in percentage to trigger the use of the optimisation.

Optimization is enabled during prediction only when use_for_optim is true. metric_names lists the metrics used for the optimization.

When the export_fitting parameter is enabled, optimization curves are exported during training.

Steps impacted

training

Example

fit_parameters: 
    resolution: 
    - locid 
    gain_threshold: 0.1 
    use_for_optim: true
    metric_names:
    - precision
    export_fitting: true

Conditional overriding of Hyper Parameters

Verteego support overriding of Hyper Parameters for a given resolution, using the following section in the config:

conditional_algorithm_parameters:
  usa:
    where:
      country: us
    alpha: 0.666
  france:
    where:
      country: fr
    alpha: 0.777

Please note that the condition expressed in the where statement must be unique for a given model resolution.