Interpreting Models

H2O Eval Studio provides interpretability of machine learning models. Model interpretations can be run either from the command line interface or using the Python API.

Currently supported models include:

Jupyter Notebooks with examples for all supported models:

Explainable Model

h2o_sonar.lib.api.models.ExplainableModel is typically used when there is a need to explain the model which is not yet supported by the library. In this case, the explainable model instance (implemented by the library user) provides model’s metadata and prediction function. Explainable model instance is model agnostic and can be used with any model type so that it can be explainable by the library.

Explainable Model Handle

h2o_sonar.lib.api.models.ExplainableModelHandle represents a remote model hosted e.g. by a Driverless AI server. For example, it is used by H2O Model Validation explainers, which use Driverless AI servers as workers to explain models. Explainable model handle string serialization (used, for example, on the command line) has the following format:

resource:connection:<connection ID>:key:<model ID>

where:

  • connection ID

    • … is a unique identifier of the Driverless AI connection specified in the H2O Eval Studio configuration.

  • model ID

    • … is a unique identifier of the model hosted by the Driverless AI server (typically UUID).

Example:

resource:connection:local-driverless-ai-server:key:7965e2ea-f898-11ed-b979-106530ed5ceb

Interpreting Models using the Command Line Interface

Use --help option to determine interpretation options:

$ h2o-sonar --help
    H2O Eval Studio Python library for Responsible AI.

    ...

    optional arguments per action and entity:

      ...

      run interpretation:
        ...
        --model          path to the serialized model or model handle.
        --model-type     model type: 'pickle' or 'mojo'
  • --model

    • Path to the model persisted on the filesystem or model handle.

  • --model-type
    • Optional model type specification (typically auto detected):
      • mojo

        • Driverless AI model saved as MOJO (.mojo file extension).

      • pickle

Example of pickled Scikit-learn model interpretation:

$ h2o-sonar run interpretation
    --dataset=dataset.csv
    --target-col=TARGET
    --results-location=./results
    --model-type=pickle
    --model=model.pkl

Example of running selected explainers to interpret the model - first find out what are the available explainers, then run selected:

$ h2o-sonar list explainers

[
    'h2o_sonar.explainers.fi_naive_shapley_explainer.NaiveShapleyMojoFeatureImportanceExplainer',
    'h2o_sonar.explainers.summary_shap_explainer.SummaryShapleyExplainer',
    'h2o_sonar.explainers.dt_surrogate_explainer.DecisionTreeSurrogateExplainer',
    'h2o_sonar.explainers.pd_ice_explainer.PdIceExplainer',
    'h2o_sonar.explainers.dia_explainer.DiaExplainer',
    'h2o_sonar.explainers.transformed_fi_shapley_explainer.ShapleyMojoTransformedFeatureImportanceExplainer'
]

$ h2o-sonar run interpretation
    --explainers=h2o_sonar.explainers.summary_shap_explainer.SummaryShapleyExplainer,h2o_sonar.explainers.pd_ice_explainer.PdIceExplainer
    --dataset=dataset.csv
    --target-col=TARGET
    --results-location=./results
    --model-type=pickle
    --model=model.pkl

Example of H2O Driverless AI model model interpretation over REST:

$ h2o-sonar run interpretation
    --dataset=dataset.csv
    --target-col=TARGET
    --results-location=./results
    --model=http://localhost:8080/model

See also Jupyter Notebook Examples:

Interpreting Models using the Python API

See Jupyter Notebook Examples:

  • Explainers Examples