Surrogate Models

Fiddler’s explainability features require a model on the backend that can generate explanations for you.

A surrogate model is an approximation of your model using gradient boosted trees (LightGBM), trained with a general, predefined set of hyperparameters. It serves as a way for Fiddler to generate approximate explanations without you having to upload your actual model artifact.

A surrogate model will be built automatically for you when you call add_model_surrogate.
You just need to provide a few pieces of information about how your model operates.

What you need to specify

• Your model’s task (regression, binary classification, etc.)
• Your model’s target column (ground truth labels)
• Your model’s output column (model predictions)
• Your model’s feature columns