The ML.GET_INSIGHTS function

This document describes the ML.GET_INSIGHTS function, which you can use to retrieve information about changes to key metrics in your multi-dimensional data from a contribution analysis model. You can use a CREATE MODEL statement to create a contribution analysis model in BigQuery.

Syntax

ML.GET_INSIGHTS(
  MODEL `PROJECT_ID.DATASET.MODEL_NAME`
)

Arguments

ML.GET_INSIGHTS takes the following arguments:

PROJECT_ID: Your project ID.
DATASET: The BigQuery dataset that contains the model.
MODEL_NAME: The name of the contribution analysis model.

Output

Some of the ML.GET_INSIGHTS output columns contain metrics that compare the values for a given segment in either the test or control dataset against the values for the population, which is all segments in the same dataset. The metric values calculated for the entire population except for the given segment are referred to as complement values.

Output for summable metric contribution analysis models

ML.GET_INSIGHTS returns the following output columns for contribution analysis models that use summable metrics, in addition to the dimension columns:

contributors: an ARRAY<STRING> value that contains the dimension values for a given segment. The other output metrics that are returned in the same row apply to the segment described by these dimensions.
metric_test: a numeric value that contains the sum of the value of the metric column in the test dataset for the given segment. The metric column is specified in the CONTRIBUTION_METRIC option of the contribution analysis model.

SUM(metric_column_name) WHERE is_test_col = TRUE
metric_control: a numeric value that contains the sum of the value of the metric column in the control dataset for the given segment. The metric column is specified in the CONTRIBUTION_METRIC option of the contribution analysis model.

SUM(metric_column_name) WHERE is_test_col = FALSE
difference: a numeric value that contains the difference between the metric_test and metric_control values:

metric_test - metric_control
relative_difference: a numeric value that contains the relative change in the segment value between the test and control datasets:

difference / metric_control
unexpected_difference: a numeric value that contains the unexpected difference between the segment's actual metric_test value and the segment's expected metric_test value, which is determined by comparing the ratio of change for this segment against the complement ratio of change. The unexpected_difference value is calculated as follows:
1. Determine the metric_test value for all segments except the given segment, referred to here as complement_test_change:
  
  complement_test_change = sum(metric_test for the population) - metric_test
2. Determine the metric_control value for all segments except the given segment, referred to here as complement_control_change:
  
  complement_control_change = sum(metric_control for the population) - metric_control
3. Determine the ratio between the complement_test_change and complement_control_change values, referred to here as complement_change_ratio:
  
  complement_change_ratio = complement_test_change / complement_control_change
4. Determine the expected metric_test value for the given segment, referred to here as expected_metric_test:
  
  expected_metric_test = metric_control * complement_change_ratio
5. Determine the unexpected_difference value:
  
  unexpected_difference = metric_test - expected_metric_test
relative_unexpected_difference: a numeric value that contains the ratio between the unexpected_difference value and the expected_metric_test value:

unexpected_difference / expected_metric_test

You can use the relative_unexpected_difference value to determine if the change to this segment is smaller than expected compared to the change in all of the other segments.
apriori_support: a numeric value that contains the apriori support value for the segment. The apriori support value is either the ratio between the metric_test value for the segment and the metric_test value for the population, or the ratio between the metric_control value for the segment and the metric_control value for the population, whichever is greater. The calculation is expressed as the following:
```
GREATEST(
  metric_test / SUM(metric_test for the population),
  metric_control / SUM(metric_control for the population)
)
```
If the apriori_support value is less than the apriori support threshold value specified in the model, then the segment is considered too small to be of interest and is excluded by the model.
contribution: a numeric value that contains the absolute value of the difference value: ABS(difference).

Insights are automatically ordered by contribution in descending order to determine the contributors associated with the largest differences in your data between the test and control sets.

Output for summable ratio metric contribution analysis models

ML.GET_INSIGHTS returns the following output columns for contribution analysis models that use summable ratio metrics, in addition to the dimension columns:

contributors: an ARRAY<STRING> value that contains the dimension values for a given segment. The other output metrics that are returned in the same row apply to the segment described by these dimensions.
metric_test: a numeric value that contains the ratio between the two metrics that you are evaluating, in the test dataset for the given metric. These two metrics are specified in the CONTRIBUTION_METRIC option of the contribution analysis model. The metric_test value is calculated as the following:

sum(numerator_metric_column_name) / sum(denominator_metric_column_name) WHERE is_test_col = TRUE
metric_control: a numeric value that contains the ratio between the two metrics that you are evaluating, in the control dataset for the given metric. These two metrics are specified in the CONTRIBUTION_METRIC option of the contribution analysis model. The metric_control value is calculated as the following:

SUM(numerator_metric_column_name) / SUM(denominator_metric_column_name) WHERE is_test_col = FALSE
metric_test_over_metric_control: a numeric value that contains the ratio between the metric_test value and the metric_control value:

metric_test / metric_control

metric_test_over_complement: a numeric value that contains the ratio between the metric_test value for this segment and the complement metric_test value:

metric_test / SUM(metric_test for the complement)

You can use the metric_test_over_complement value to compare the size of this segment to the size the other segments.

For example, consider the following table of test data:

dim1	dim2	dim3	metric_a	metric_b
1	10	20	50	100
1	15	30	75	200
5	20	40	1	10

Assume that the CONTRIBUTION_METRIC value is SUM(metric_a) / SUM(metric_b). Using the data in the preceding table, the metric_a value for the population is 126, while the metric_b value for the population is 310. The metric_test_over_complement value for the segment in the first row of the table is calculated as the following:

(50/100)/((75+1)/(200+10)) = .5/(76/210) = 1.38

This metric_test_over_complement value indicates that the size of this segment is larger than the size of all of the other segments combined. Alternatively, the metric_test_over_complement value for the segment in the third row of table is calculated as the following:

(1/10)/((50+75)/(100+200)) = .1/(125/300) = 0.24

This metric_test_over_complement value indicates that the size of this segment is smaller than the combined size of the rest of the segments.

metric_control_over_complement: a numeric value that contains the ratio between the metric_control value for this segment and the complement metric_control value:

metric_control / sum(metric_control for the complement)

You can use the metric_control_over_complement value to compare the size of this segment to the size of the other segments.
aumann_shapley_attribution: a numeric value that contains the Aumann-Shapley value for this segment. The Aumann-Shapley value measures the contribution of the segment ratio relative to the population ratio. You can use the Aumann-Shapley value to determine how much a feature contributes to the prediction value. In the context of contribution analysis, BigQuery ML uses the Aumann-Shapley value to measure the attribution of the segment relative to the population. When calculating this measurement, the service considers the segment ratio changes and the complement population changes between the test and control datasets.
apriori_support: a numeric value that contains the apriori support value for the segment. The apriori support value is calculated using the numerator column specified in the model's CONTRIBUTION_METRIC option:

numerator column value for the given segment / SUM(numerator column value for the population)

If the apriori_support value is less than the apriori support threshold value specified in the model, then the segment is considered too small to be of interest and is excluded by the model.
contribution: a numeric value that contains the absolute value of the aumann_shapley_attribution:

ABS(aumann_shapley_attribution)

Insights are automatically ordered by contribution in descending order to determine the contributors associated with the largest differences in your data between the test and control sets.

Output for summable by category metric contribution analysis models

ML.GET_INSIGHTS returns the following output columns for contribution analysis models that use summable category metrics, in addition to the dimension columns:

contributors: an ARRAY<STRING> value that contains the dimension values for a given segment. The other output metrics that are returned in the same row apply to the segment described by these dimensions.
metric_test: a numeric value that contains the ratio between the sum of the metric column and the number of distinct values of the count distinct column in the test dataset for a given segment:

SUM(sum_column_name) / COUNT(DISTINCT categorical_column_name) WHERE is_test_col = TRUE

The metric and count distinct columns are specified in the CONTRIBUTION_METRIC option of the contribution analysis model.
metric_control: a numeric value that contains the ratio between the sum of the metric column and the number of distinct values of the count distinct column in the control dataset for a given segment:

SUM(sum_column_name) / COUNT(DISTINCT categorical_column_name) WHERE is_test_col = FALSE

The metric and categorical columns are specified in the CONTRIBUTION_METRIC option of the contribution analysis model.
difference: a numeric value that contains the difference between the metric_test and metric_control values:

metric_test - metric_control.
relative_difference: a numeric value that contains the relative change in the segment value between the test and control datasets:

difference/metric_control
metric_test_over_population: a numeric value that contains the ratio between the metric_test value for this segment and the metric_test value for the population:

metric_test / (metric_test for the population)

You can use the metric_test_over_population value to compare the size of the segment to the overall size of the test dataset.
metric_control_over_population: a numeric value that contains the ratio between the metric_control value for this segment and the metric_control value for the population:

metric_control / (metric_control for the population)

You can use the metric_control_over_population value to compare the size of the segment to the overall size of the control dataset.
apriori_support: a numeric value that contains the apriori support value for the segment. To calculate apriori support, the sum_metric_column is used to compute the segment size relative to the population for both the test and control datasets and apriori_support is selected as the greater of the two values. The calculation is expressed as the following:
```
GREATEST(
  SUM(sum_column_name test) / SUM(sum_column_name test for the population),
  SUM(sum_column_name control) / SUM(sum_column_name control for the population)
)
```
If the apriori_support value is less than the apriori support threshold value specified in the model, then the segment is considered too small to be of interest and is excluded by the model.
contribution: a numeric value that contains the absolute value of the difference, calculated as ABS(difference).

Insights are automatically ordered by contribution in descending order to quickly determine the contributors associated with the largest differences in your data between the test and control sets.

What's next

Get data insights from a contribution analysis model.