Model Agnostics uses the ofs_model_xray package.Peek file located inside the black box to better understand the model. Model explanations add important criteria for model selection. Compare models in terms of interpretability when working with model-agnostic explanations, because the same method can be used for any type of model.
The following three types of explanations are supported:
· Global Explanations: This explanation talks about the general behavior of the model.
§ Feature Importance: This explanation measures how much a model relies on a feature to make a prediction, regardless of the performance accuracy. The most important features are calculated on the training dataset. These features are the ones that affect the performance the most.
· Local Explanations: This explanation is used for specific model predictions.
§ Feature Contribution: For any given prediction or observation, this gives the contribution of each of the features. It is computed using the Shapley Sampling algorithm and is calculated on the validation dataset.
· A mix of Global and Local Explanations: This explanation is used for specific model explanations.
§ Model Response (PDP/ICE): This explanation indicates how the model responds (mean prediction) as the feature value changes. The model response measures the partial dependence of the model on a feature, that is, the change in the average predicted value as specified features change over their marginal distribution. In addition to the model mean response, conditional responses for those individual observations are also computed for a random sample of observations. This is calculated on the validation dataset.
Use the aif.explain_model python API to perform model agnostics or explanations.
The following are the Inputs:
· Technique: Model agnostic technique.
§ Permuted Feature Importance
§ Model Response
§ Feature Contribution
For example, Model Response.
· model_id: Input model ID. For the exact model ID, see the Model column in the AUC Summary.
For example, XGB1.
· entity_id : Input entity ID
For example, CUST_ID_1234.
The output is the model agnostic plots. For more information on model agnostics, see the following links:
· http://www.jmlr.org/papers/volume11/strumbelj10a/strumbelj10a.pdf
· https://www.stat.berkeley.edu/~breiman/randomforest2001.pdf