arthurai.explainability.arthur_explainer.ArthurExplainer

class arthurai.explainability.arthur_explainer.ArthurExplainer(model_type, model_input_type, predicted_attributes, predict_func, data, label_mapping, categorical_features=None, enable_shap=True, shap_training_samples=20, enable_lime=True, text_delimiter='\\\\W+', load_image_func=None, predictions_df=None)

Bases: object

Class for explaining inferences with a variety of libs

Methods

build_image_explainers	rtype None
build_nlp_explainers	rtype None
build_tabular_explainers	rtype None
decode_labels	Decodes into strings
encode_labels	Encodes labels into integers
explain_image	rtype List[Dict[str, Any]]
explain_image_lime	rtype List[Dict[str, Any]]
explain_nlp	rtype List[List[List[List[float]]]]
explain_nlp_lime	rtype List[List[List[List[float]]]]
explain_tabular	rtype List[List[List[float]]]
explain_tabular_lime	Calculate lime values.
explain_tabular_shap	Calculate SHAP values.
format_lime_image_results	Format LIME results for images
format_lime_results	LIME's .as_map() method returns scores looking like the following:
format_lime_results_for_one_class	This method unpacks the results for a single classes's explanation:
format_lime_text_results	Format LIME results for Text models
predict_tabular	Takes in customer provided data and returns the prediction values as numpy array (both regression and classification)
predict_wrapper	Wraps user supplied predict function.
shap_expected_values	Returns expected values from the shap explainer for this model input type

Attributes

LIME
SHAP
num_predicted_attributes	rtype int

decode_labels(feature_vector)

Decodes into strings

Return type

Iterable

encode_labels(feature_vector)

Encodes labels into integers

Return type

Iterable

explain_tabular_lime(raw_feature_vectors, **kwargs)

Calculate lime values. For one or more inferences.

Parameters

• raw_feature_vectors (List) – An iterable of one or more feature vectors to explain.

• kwargs – Additional args to pass to LIME.

Return type

List[List[List[float]]]

Returns

A List of explanations for each feature vector. A single explanation is a list of lists that contain the explanation for each model output by input feature. In this example, the model has 4 features and 3 output classes:

[
    [
        [-0.011, 0.009, -0.128, -0.161],
        [-0.009, 0.003, -0.091, -0.065],
        [0.02, -0.013, 0.220, 0.226]
    ]
]

explain_tabular_shap(raw_feature_vectors, **kwargs)

Calculate SHAP values. For one or more inferences.

Parameters

• raw_feature_vectors (List) – An iterable of one or more feature vectors to explain.

• kwargs – Additional args to pass to SHAP.

Return type

List[List[List[float]]]

Returns

A List of explanations for each feature vector. A single explanation is a list of lists that contain the explanation for each model output by input feature. In this example, the model has 4 features and 3 output classes.

[
    [-0.011, 0.009, -0.128, -0.161],
    [-0.009, 0.003, -0.091, -0.065],
    [0.02, -0.013, 0.220, 0.226]
]

static format_lime_image_results(results)

Format LIME results for images

Return type

List[Dict[str, Any]]

Returns

formatted output in the following format:

{
    "lime_segment_mask": [
        [1, 1, 1, 3, 3, 1, 0],
        [1, 1, 1, 3, 3, 1, 0],
        [1, 1, 1, 3, 3, 1, 0],
        [2, 2, 2, 3, 3, 1, 0]
    ],
    "lime_region_mapping": {
        "0": [[0, 0.23], [1, 0.3], [2, 0.001], [3, -0.5]]
    }
}

format_lime_results(results)

LIME’s .as_map() method returns scores looking like the following:

{
   0:[
      (3, -0.1614709564346036),
      (2, -0.12871035938256958),
      (0, -0.011225126011010974),
      (1, 0.009948039509525414)
   ],
   1:[
      (2, -0.0914379318506283),
      (3, -0.06513690080871648),
      (0, -0.009164069675758498),
      (1, 0.003622751782692439)
   ],
   2:[
      (3, 0.2266078572433201),
      (2, 0.22014829123319787),
      (0, 0.02038919568676947),
      (1, -0.013570791292217853)
   ]
}

Where the keys are the class labels, and the values are lists of feature index, score tuples for that class.

This method unpacks the result to a list of lists for EACH explanation:

[
    [-0.011, 0.009, -0.128, -0.161],
    [-0.009, 0.003, -0.091, -0.065],
    [0.02, -0.013, 0.220, 0.226]
]

For regression, this will return a list of lists for one class per inference:

[
    [-0.011, 0.009, -0.128, -0.161]
]

Return type

List[List[List[float]]]

static format_lime_results_for_one_class(feature_index_score_pairs)

This method unpacks the results for a single classes’s explanation:

[
    (3, -0.1614709564346036),
    (2, -0.12871035938256958),
    (0, -0.011225126011010974),
    (1, 0.009948039509525414)
]

Return type

List[float]

Returns

unpacked lime results for a single class

[-0.011, 0.009, -0.128, -0.161]

format_lime_text_results(results)

Format LIME results for Text models

LIME’s .as_map() method returns scores looking like the following:

{0: [('I', -0.1614709564346036), ('a', -0.12871035938256958), ('am', -0.011225126011010974), ('developer', 0.009948039509525414)],
 1: [('am', -0.0914379318506283), ('a', -0.06513690080871648), ('I', -0.009164069675758498), ('developer', 0.003622751782692439)],
 2: [('developer', 0.2266078572433201), ('am', 0.22014829123319787), ('a', 0.02038919568676947), ('I', -0.013570791292217853)]}

Where the keys are the class labels, and the values are lists of word, score tuples for that class.

This method unpacks the result to a list of lists for EACH explanation:

[
    [['I', -0.011], ['am', 0.009], ['a', -0.128], ['developer', -0.161]],
    [['I', -0.02], ['am', 0.007], ['a', -0.8], ['developer', 0.11]],
    [['I', -0.03], ['am', 0.89], ['a', -0.28], ['developer', -0.16]]
]

For regression, this will return a list of lists for one class per inference:

[
    [['I', -0.011], ['am', 0.009], ['a', -0.128], ['developer', -0.161]]
]

Return type

List[List[List[List[float]]]]

Returns

formatted results in the following format

[
    [['I', -0.011], ['am', 0.009], ['a', -0.128], ['developer', -0.161]],
    [['I', -0.02], ['am', 0.007], ['a', -0.8], ['developer', 0.11]],
    [['I', -0.03], ['am', 0.89], ['a', -0.28], ['developer', -0.16]]
]

predict_tabular(input_feature_vectors)

Takes in customer provided data and returns the prediction values as numpy array (both regression and classification)

Return type

ndarray

predict_wrapper(encoded_feature_vectors)

Wraps user supplied predict function. Handles decoding back into original format. For binary classifiers, checks the output size and reshapes to two dimensions if necessary

Return type

ndarray

shap_expected_values()

Returns expected values from the shap explainer for this model input type

Return type

List[float]