Explanation API and feature importance for GBM

[jppgks]

Introduces abstract Explainer class with the following method for explaining model predictions:

def explain(self, **kwargs) -> Tuple[List[Explanation], List[float]]:
    """Explain the model's predictions.

    # Return

    :return: (Tuple[List[Explanation], List[float]]) `(explanations, expected_values)`
        `explanations`: (List[Explanation]) A list of explanations, one for each row in the input data. Each
        explanation contains the feature attributions for each label in the target feature's vocab.

        `expected_values`: (List[float]) of length [output feature cardinality] Expected value for each label in
        the target feature's vocab.
    """

Explanation is a dataclass for storing explanations for a single row of input data.

Additionally:

• refactors Integrated Gradients as an implementation of the Explainer base class
• adds GBMExplainer for explaining GBM models using LightGBM feature importance

[github-actions]

Unit Test Results

         6 files           6 suites   3h 9m 21s ⏱️
  3 428 tests   3 350 ✔️   78 💤 0 ❌
10 284 runs  10 027 ✔️ 257 💤 0 ❌

Results for commit 14306a5.

♻️ This comment has been updated with latest results.

[jppgks]

@tgaddair what should the expected values be for GBM?

[jppgks]

From Travis: value of 0 OK for now, can look into this again depending on future visualizations

[dantreiman]

Code looks great, very positive improvement toludwig.explain. Made a few optional refactoring suggestions

[dantreiman]

LGTM!