The ML.EXPLAIN_PREDICT function

This document describes the ML.EXPLAIN_PREDICT function, which lets you generate a predicted value and a set of feature attributions for each instance of the input data. Feature attributions indicate how much each feature in your model contributed to the final prediction for each given instance. ML.EXPLAIN_PREDICT is essentially an extended version of ML.PREDICT.

Syntax

ML.EXPLAIN_PREDICT(
  MODEL `project_id.dataset.model_name`,
  { TABLE `project_id.dataset.table` | (query_statement) },
  STRUCT(
  [number_of_output_tokens AS max_output_tokens]
  [, top_k AS top_k_features]
  [, threshold AS threshold]
  [, integrated_gradients_num_steps AS integrated_gradients_num_steps]
  [, approx_feature_contrib AS approx_feature_contrib])
)

Arguments

ML.EXPLAIN_PREDICT takes the following arguments:

  • project_id: Your project ID.
  • dataset: The BigQuery dataset that contains the model.
  • model: The name of the model.
  • table: The name of the input table that contains the data to be evaluated.

    If table is specified, the input column names in the table must match the column names in the model, and their types should be compatible according to BigQuery implicit coercion rules.

    If there are unused columns from the table, they are passed through to the output columns.

  • query_statement: The GoogleSQL query that is used to generate the evaluation data. For the supported SQL syntax for the query_statement clause in GoogleSQL, see Query syntax.

    If query_statement is specified, the input column names from the query must match the column names in the model, and their types should be compatible according to BigQuery implicit coercion rules.

    If there are unused columns from the table, they are passed through to the output columns.

    If you used the TRANSFORM clause in the CREATE MODEL statement that created the model, then only the input columns present in the TRANSFORM clause can appear in query_statement.

  • top_k_features: an INT64 value that specifies how many top feature attribution pairs are generated for each row of input data. The features are ranked by the absolute values of their attributions.

    By default, top_k_features is set to 5. If its value is greater than the number of features in the training data, the attributions of all features are returned.

  • threshold: a FLOAT64 value that specifies the cutoff between the two labels for binary classification models. Predictions above the threshold are positive predictions. Predictions below the threshold are negative predictions. Feature attributions are returned only for the predicted label.

    The threshold value must be between 0.0 and 1.0. The default value is 0.5.

  • integrated_gradients_num_steps: an INT64 value that specifies the number of steps to sample between the example being explained and its baseline. This value is used to approximate the integral in integrated gradients attribution methods. Increasing the value improves the precision of feature attributions, but can be slower and more computationally expensive.

    This option only applies to deep neural network (DNN) models, which use integrated gradients attribution methods. The default value is 15.

  • approx_feature_contrib: a BOOL value that indicates whether to use approximate feature contribution method in XGBoost model explanation. This option applies only to boosted tree and random forest models.

    This capability is provided by the XGBoost library; BigQuery ML only passes this option through to it. For more information, see Package 'xgboost' and search for approxcontrib.

    The default value is FALSE.

Output

ML.EXPLAIN_PREDICT returns the following columns in addition to any passthrough columns:

  • predicted_<label_column_name>: a STRING value that contains either the predicted value of the label for regression models or the predicted label class for classification models.
  • probability: a FLOAT64 value that contains the probability of the predicted label class. This column is only present for classification models.
  • top_feature_attributions: An ARRAY<STRUCT> value that contains the attributions of the top k features to the final prediction:
    • top_feature_attributions.feature: a STRING value that contains the feature name.
    • top_feature_attributions.attribution: a FLOAT64 value that contains the attribution of the feature to the final prediction.
  • baseline_prediction_value: a FLOAT64 value that contains one of the following:
    • For linear models, the baseline_prediction_value value is the intercept of the model.
    • For DNN models, the baseline_prediction_value value is the mean across all numerical features and NULL for other types of features.
    • For boosted tree and random forest models, the baseline_prediction_value value is equal to the bias term, which is the expected output of the model over the training dataset. See Tree SHAP documentation for more information.
  • prediction_value: The raw prediction value.
    • For regression models, this is a FLOAT64 value that contains the value of the column identified by predicted_<label_column_name>.
    • For classification models, this is an INT or STRING value that contains the logit value (also called log-odds) for the predicted class. The predicted class probabilities are obtained by applying the softmax transformation to the logit values.
  • approximation_error:

    • Exact attribution methods like Tree SHAP are defined a