alibi.explainers.anchors.anchor_tabular module

class alibi.explainers.anchors.anchor_tabular.AnchorTabular(predictor, feature_names, categorical_names=None, dtype=<class 'numpy.float32'>, ohe=False, seed=None)[source]

Bases: Explainer, FitMixin

__init__(predictor, feature_names, categorical_names=None, dtype=<class 'numpy.float32'>, ohe=False, seed=None)[source]
  • predictor (Callable[[ndarray], ndarray]) – A callable that takes a numpy array of N data points as inputs and returns N outputs.

  • feature_names (List[str]) – List with feature names.

  • categorical_names (Optional[Dict[int, List[str]]]) – Dictionary where keys are feature columns and values are the categories for the feature.

  • dtype (Type[generic]) – A numpy scalar type that corresponds to the type of input array expected by predictor. This may be used to construct arrays of the given type to be passed through the predictor. For most use cases this argument should have no effect, but it is exposed for use with predictors that would break when called with an array of unsupported type.

  • ohe (bool) – Whether the categorical variables are one-hot encoded (OHE) or not. If not OHE, they are assumed to have ordinal encodings.

  • seed (Optional[int]) – Used to set the random number generator for repeatability purposes.


Add feature names to explanation dictionary.


explanation (dict) – Dict with anchors and additional metadata.

Return type:


explain(X, threshold=0.95, delta=0.1, tau=0.15, batch_size=100, coverage_samples=10000, beam_size=1, stop_on_first=False, max_anchor_size=None, min_samples_start=100, n_covered_ex=10, binary_cache_size=10000, cache_margin=1000, verbose=False, verbose_every=1, **kwargs)[source]

Explain prediction made by classifier on instance X.

  • X (ndarray) – Instance to be explained.

  • threshold (float) – Minimum anchor precision threshold. The algorithm tries to find an anchor that maximizes the coverage under precision constraint. The precision constraint is formally defined as \(P(prec(A) \ge t) \ge 1 - \delta\), where \(A\) is an anchor, \(t\) is the threshold parameter, \(\delta\) is the delta parameter, and \(prec(\cdot)\) denotes the precision of an anchor. In other words, we are seeking for an anchor having its precision greater or equal than the given threshold with a confidence of (1 - delta). A higher value guarantees that the anchors are faithful to the model, but also leads to more computation time. Note that there are cases in which the precision constraint cannot be satisfied due to the quantile-based discretisation of the numerical features. If that is the case, the best (i.e. highest coverage) non-eligible anchor is returned.

  • delta (float) – Significance threshold. 1 - delta represents the confidence threshold for the anchor precision (see threshold) and the selection of the best anchor candidate in each iteration (see tau).

  • tau (float) – Multi-armed bandit parameter used to select candidate anchors in each iteration. The multi-armed bandit algorithm tries to find within a tolerance tau the most promising (i.e. according to the precision) beam_size candidate anchor(s) from a list of proposed anchors. Formally, when the beam_size=1, the multi-armed bandit algorithm seeks to find an anchor \(A\) such that \(P(prec(A) \ge prec(A^\star) - \tau) \ge 1 - \delta\), where \(A^\star\) is the anchor with the highest true precision (which we don’t know), \(\tau\) is the tau parameter, \(\delta\) is the delta parameter, and \(prec(\cdot)\) denotes the precision of an anchor. In other words, in each iteration, the algorithm returns with a probability of at least 1 - delta an anchor \(A\) with a precision within an error tolerance of tau from the precision of the highest true precision anchor \(A^\star\). A bigger value for tau means faster convergence but also looser anchor conditions.

  • batch_size (int) – Batch size used for sampling. The Anchor algorithm will query the black-box model in batches of size batch_size. A larger batch_size gives more confidence in the anchor, again at the expense of computation time since it involves more model prediction calls.

  • coverage_samples (int) – Number of samples used to estimate coverage from during result search.

  • beam_size (int) – Number of candidate anchors selected by the multi-armed bandit algorithm in each iteration from a list of proposed anchors. A bigger beam width can lead to a better overall anchor (i.e. prevents the algorithm of getting stuck in a local maximum) at the expense of more computation time.

  • stop_on_first (bool) – If True, the beam search algorithm will return the first anchor that has satisfies the probability constraint.

  • max_anchor_size (Optional[int]) – Maximum number of features in result.

  • min_samples_start (int) – Min number of initial samples.

  • n_covered_ex (int) – How many examples where anchors apply to store for each anchor sampled during search (both examples where prediction on samples agrees/disagrees with desired_label are stored).

  • binary_cache_size (int) – The result search pre-allocates binary_cache_size batches for storing the binary arrays returned during sampling.

  • cache_margin (int) – When only max(cache_margin, batch_size) positions in the binary cache remain empty, a new cache of the same size is pre-allocated to continue buffering samples.

  • verbose (bool) – Display updates during the anchor search iterations.

  • verbose_every (int) – Frequency of displayed iterations during anchor search process.

Return type:



explanationExplanation object containing the result explaining the instance with additional metadata as attributes. See usage at AnchorTabular examples for details.


alibi.exceptions.NotFittedError – If fit has not been called prior to calling explain.

fit(train_data, disc_perc=(25, 50, 75), **kwargs)[source]

Fit discretizer to train data to bin numerical features into ordered bins and compute statistics for numerical features. Create a mapping between the bin numbers of each discretised numerical feature and the row id in the training set where it occurs.

  • train_data (ndarray) – Representative sample from the training data.

  • disc_perc (Tuple[Union[int, float], ...]) – List with percentiles (int) used for discretization.

Return type:


instance_label: int

The label of the instance to be explained.

meta: dict

Object metadata.

property predictor: Callable | None
Return type:



Resets the predictor function.


predictor (Callable) – New predictor function.

Return type:


samplers: list
class alibi.explainers.anchors.anchor_tabular.TabularSampler(predictor, disc_perc, numerical_features, categorical_features, feature_names, feature_values, n_covered_ex=10, seed=None)[source]

Bases: object

A sampler that uses an underlying training set to draw records that have a subset of features with values specified in an instance to be explained, X.

__call__(anchor, num_samples, compute_labels=True)[source]

Obtain perturbed records by drawing samples from training data that contain the categorical labels and discretized numerical features and replacing the remainder of the record with arbitrary values.

  • anchor (Tuple[int, tuple]) – The integer represents the order of the result in a request array. The tuple contains encoded feature indices.

  • num_samples (int) – Number of samples used when sampling from training set.

  • compute_labels – If True, an array of comparisons between predictions on perturbed samples and instance to be explained is returned.

Return type:

Union[List[Union[ndarray, float, int]], List[ndarray]]


  • If compute_labels=True, a list containing the following is returned –

    • covered_true - perturbed examples where the anchor applies and the model prediction on perturbation is the same as the instance prediction.

    • covered_false - perturbed examples where the anchor applies and the model prediction is NOT the same as the instance prediction.

    • labels - num_samples ints indicating whether the prediction on the perturbed sample matches (1) the label of the instance to be explained or not (0).

    • data - Sampled data where ordinal features are binned (1 if in bin, 0 otherwise).

    • coverage - the coverage of the anchor.

    • anchor[0] - position of anchor in the batch request.

  • Otherwise, a list containing the data matrix only is returned.

__init__(predictor, disc_perc, numerical_features, categorical_features, feature_names, feature_values, n_covered_ex=10, seed=None)[source]
  • predictor (Callable) – A callable that takes a tensor of N data points as inputs and returns N outputs.

  • disc_perc (Tuple[Union[int, float], ...]) – Percentiles used for numerical feature discretisation.

  • numerical_features (List[int]) – Numerical features column IDs.

  • categorical_features (List[int]) – Categorical features column IDs.

  • feature_names (list) – Feature names.

  • feature_values (dict) – Key: categorical feature column ID, value: values for the feature.

  • n_covered_ex (int) – For each result, a number of samples where the prediction agrees/disagrees with the prediction on instance to be explained are stored.

  • seed (Optional[int]) – If set, fixes the random number sequence.


An encoding of the feature IDs is created by assigning each bin of a discretized numerical variable and each categorical variable a unique index. For a dataset containing, e.g., a numerical variable with 5 bins and 3 categorical variables, indices 0 - 4 represent bins of the numerical variable whereas indices 5, 6, 7 represent the encoded indices of the categorical variables (but see note for caviats). The encoding is necessary so that the different ranges of the numerical variable can be sampled during result construction. Note that the encoded indices represent the predicates used during the anchor construction process (i.e., and anchor is a collection of encoded indices.


X (ndarray) – Instance to be explained.

Return type:



A list containing three dictionaries, whose keys are encoded feature IDs

  • cat_lookup - maps categorical variables to their value in X.

  • ord_lookup - maps discretized numerical variables to the bins they can be sampled from given X.

  • enc2feat_idx - maps the encoded IDs to the original (training set) feature column IDs.


Each continuous variable has n_bins - 1 corresponding entries in ord_lookup.


Compute the agreement between a classifier prediction on an instance to be explained and the prediction on a set of samples which have a subset of features fixed to specific values.


samples (ndarray) – Samples whose labels are to be compared with the instance label.

Return type:



An array of integers indicating whether the prediction was the same as the instance label.

deferred_init(train_data, d_train_data)[source]

Initialise the tabular sampler object with data, discretizer, feature statistics and build an index from feature values and bins to database rows for each feature.

  • train_data (Union[ndarray, Any]) – Data from which samples are drawn. Can be a numpy array or a ray future.

  • d_train_data (Union[ndarray, Any]) – Discretized version for training data. Can be a numpy array or a ray future.

Return type:



An initialised sampler.


Given an anchor, this function finds the row indices in the training set where the feature has the same value as the feature in the instance to be explained (for ordinal variables, the row indices are those of rows which contain records with feature values in the same bin). The algorithm uses both the feature encoded ids in anchor and the feature ids in the input data set. The two are mapped by self.enc2feat_idx.


anchor (tuple) – The anchor for which the training set row indices are to be retrieved. The ints represent encoded feature ids.

Return type:

Tuple[Dict[int, Set[int]], Dict[int, Any], List[Tuple[int, str, Union[Any, int]]]]


  • allowed_bins – Maps original feature ids to the bins that the feature should be sampled from given the input anchor.

  • allowed_rows – Maps original feature ids to the training set rows where these features have the same value as the anchor.

  • unk_feat_values – When a categorical variable with the specified value/discretized variable in the specified bin is not found in the training set, a tuple is added to unk_feat_values to indicate the original feature id, its type ('c' = categorical, 'o' = discretized continuous) and the value/bin it should be sampled from.

handle_unk_features(allowed_bins, num_samples, samples, unk_feature_values)[source]

Replaces unknown feature values with defaults. For categorical variables, the replacement value is the same as the value of the unknown feature. For continuous variables, a value is sampled uniformly at random from the feature range.

Return type:


instance_label: int

The label of the instance to be explained.

perturbation(anchor, num_samples)[source]

Implements functionality described in alibi.explainers.anchors.anchor_tabular.TabularSampler.__call__().

  • anchor (tuple) – Each int is an encoded feature id.

  • num_samples (int) – Number of samples.

Return type:

Tuple[ndarray, ndarray, float]


  • samples – Sampled data from training set.

  • d_samples – Like samples, but continuous data is converted to ordinal discrete data (binned).

  • coverage – The coverage of the result in the training data.

replace_features(samples, allowed_rows, uniq_feat_ids, partial_anchor_rows, nb_partial_anchors, num_samples)[source]

The method creates perturbed samples by first replacing all partial anchors with partial anchors drawn from the training set. Then remainder of the features are then replaced with random values drawn from the same bin for discretized continuous features and same value for categorical features.

  • samples (ndarray) – Randomly drawn samples, where the anchor does not apply.

  • allowed_rows (Dict[int, Any]) – Maps feature ids to the rows indices in training set where the feature has same value as instance (cat.) or is in the same bin.

  • uniq_feat_ids (List[int]) – Multiple encoded features in the anchor can map to the same original feature id. Unique features in the anchor. This is the list of unique original features id in the anchor.

  • partial_anchor_rows (List[ndarray]) – The rows in the training set where each partial anchor applies. Last entry is an array of row indices where the entire anchor applies.

  • nb_partial_anchors (ndarray) – The number of training records which contain each partial anchor.

  • num_samples (int) – Number of perturbed samples to be returned.

Return type:



Sets the sampler label. Necessary for setting the remote sampling process state during explain call.


X (ndarray) – Instance to be explained.

Return type:



Set the number of examples to be saved for each result and partial result during search process. The same number of examples is saved in the case where the predictions on perturbed samples and original instance agree or disagree.


n_covered (int) – Number of examples to be saved.

Return type: