Breyta

Deila með

Generate a Responsible AI insights with YAML and Python

  • Grein

APPLIES TO: Azure CLI ml extension v2 (current) Python SDK azure-ai-ml v2 (current)

You can generate a Responsible AI dashboard and scorecard via a pipeline job by using Responsible AI components. There are six core components for creating Responsible AI dashboards, along with a couple of helper components. Here's a sample experiment graph:

Screenshot of a sample experiment graph.

Responsible AI components

The core components for constructing the Responsible AI dashboard in Azure Machine Learning are:

  • RAI Insights dashboard constructor
  • The tool components:
    • Add Explanation to RAI Insights dashboard
    • Add Causal to RAI Insights dashboard
    • Add Counterfactuals to RAI Insights dashboard
    • Add Error Analysis to RAI Insights dashboard
    • Gather RAI Insights dashboard
    • Gather RAI Insights score card

The RAI Insights dashboard constructor and Gather RAI Insights dashboard components are always required, plus at least one of the tool components. However, it isn't necessary to use all the tools in every Responsible AI dashboard.

In the following sections are specifications of the Responsible AI components and examples of code snippets in YAML and Python.

Important

Items marked (preview) in this article are currently in public preview. The preview version is provided without a service level agreement, and it's not recommended for production workloads. Certain features might not be supported or might have constrained capabilities. For more information, see Supplemental Terms of Use for Microsoft Azure Previews.

Limitations

The current set of components have many limitations on their use:

  • All models must be registered in Azure Machine Learning in MLflow format with a sklearn (scikit-learn) flavor.
  • The models must be loadable in the component environment.
  • The models must be pickleable.
  • The models must be supplied to the Responsible AI components by using the Fetch Registered Model component, which we provide.
  • The dataset inputs must be in mltable format.
  • A model must be supplied even if only a causal analysis of the data is performed. You can use the DummyClassifier and DummyRegressor estimators from scikit-learn for this purpose.

RAI Insights dashboard constructor

This component has three input ports:

  • The machine learning model
  • The training dataset
  • The test dataset

To generate model-debugging insights with components such as error analysis and Model explanations, use the training and test dataset that you used when you trained your model. For components like causal analysis, which doesn't require a model, you use the training dataset to train the causal model to generate the causal insights. You use the test dataset to populate your Responsible AI dashboard visualizations.

The easiest way to supply the model is to register the input model and reference the same model in the model input port of RAI Insight Constructor component, which we discuss later in this article.

Note

Currently, only models in MLflow format and with a sklearn flavor are supported.

The two datasets should be in mltable format. The training and test datasets provided don't have to be the same datasets that are used in training the model, but they can be the same. By default, for performance reasons, the test dataset is restricted to 5,000 rows of the visualization UI.

The constructor component also accepts the following parameters:

Parameter name Description Type
title Brief description of the dashboard. String
task_type Specifies whether the model is for classification, regression, or forecasting. String, classification, regression, or forecasting
target_column_name The name of the column in the input datasets, which the model is trying to predict. String
maximum_rows_for_test_dataset The maximum number of rows allowed in the test dataset, for performance reasons. Integer, defaults to 5,000
categorical_column_names The columns in the datasets, which represent categorical data. Optional list of strings1
classes The full list of class labels in the training dataset. Optional list of strings1
feature_metadata Specifies additional information the dashboard might need depending on task type. For forecasting, this includes specifying which column is the datetime column and which column is the time_series_id column. For vision, this might include mean pixel value or location data of an image. Optional list of strings1
use_model_dependency Specifies if the model requires a separate docker container to be served in due to conflicting dependencies with the RAI dashboard. For forecasting, this must be enabled. Typically for other scenarios this isn't enabled. Boolean

1 The lists should be supplied as a single JSON-encoded string for categorical_column_names, classes, feature_metadata inputs.

The constructor component has a single output named rai_insights_dashboard. This is an empty dashboard, which the individual tool components operate on. All the results are assembled by the Gather RAI Insights dashboard component at the end.

 create_rai_job: 

    type: command 
    component: azureml://registries/azureml/components/microsoft_azureml_rai_tabular_insight_constructor/versions/<get current version>
    inputs: 
      title: From YAML snippet 
      task_type: regression
      type: mlflow_model
      path: azureml:<registered_model_name>:<registered model version> 
      train_dataset: ${{parent.inputs.my_training_data}} 
      test_dataset: ${{parent.inputs.my_test_data}} 
      target_column_name: ${{parent.inputs.target_column_name}} 
      categorical_column_names: '["location", "style", "job title", "OS", "Employer", "IDE", "Programming language"]'

Add Causal to RAI Insights dashboard

This component performs a causal analysis on the supplied datasets. It has a single input port, which accepts the output of the RAI Insights dashboard constructor. It also accepts the following parameters:

Parameter name Description Type            
treatment_features A list of feature names in the datasets, which are potentially "treatable" to obtain different outcomes. List of strings2.
heterogeneity_features A list of feature names in the datasets, which might affect how the "treatable" features behave. By default, all features will be considered. Optional list of strings2.
nuisance_model The model used to estimate the outcome of changing the treatment features. Optional string. Must be linear or AutoML, defaulting to linear.
heterogeneity_model The model used to estimate the effect of the heterogeneity features on the outcome. Optional string. Must be linear or forest, defaulting to linear.
alpha Confidence level of confidence intervals. Optional floating point number, defaults to 0.05.
upper_bound_on_cat_expansion The maximum expansion of categorical features. Optional integer, defaults to 50.
treatment_cost The cost of the treatments. If 0, all treatments will have zero cost. If a list is passed, each element is applied to one of the treatment_features.

Each element can be a scalar value to indicate a constant cost of applying that treatment or an array indicating the cost for each sample. If the treatment is a discrete treatment, the array for that feature should be two dimensional, with the first dimension representing samples and the second representing the difference in cost between the non-default values and the default value.		 Optional integer or list2.
min_tree_leaf_samples The minimum number of samples per leaf in the policy tree. Optional integer, defaults to 2.
max_tree_depth The maximum depth of the policy tree. Optional integer, defaults to 2.
skip_cat_limit_checks By default, categorical features need to have several instances of each category in order for a model to be fit robustly. Setting this to True will skip these checks. Optional Boolean, defaults to False.
categories The categories to use for the categorical columns. If auto, the categories will be inferred for all categorical columns. Otherwise, this argument should have as many entries as there are categorical columns.

Each entry should be either auto to infer the values for that column or the list of values for the column. If explicit values are provided, the first value is treated as the "control" value for that column against which other values are compared.		 Optional, auto or list2.
n_jobs The degree of parallelism to use. Optional integer, defaults to 1.
verbose Expresses whether to provide detailed output during the computation. Optional integer, defaults to 1.
random_state Seed for the pseudorandom number generator (PRNG). Optional integer.

2 For the list parameters: Several of the parameters accept lists of other types (strings, numbers, even other lists). To pass these into the component, they must first be JSON-encoded into a single string.

This component has a single output port, which can be connected to one of the insight_[n] input ports of the Gather RAI Insights Dashboard component.

  causal_01: 
    type: command 
    component: azureml://registries/azureml/components/microsoft_azureml_rai_tabular_causal/versions/<version>
    inputs: 
      rai_insights_dashboard: ${{parent.jobs.create_rai_job.outputs.rai_insights_dashboard}} 
      treatment_features: `["Number of GitHub repos contributed to", "YOE"]'

Add Counterfactuals to RAI Insights dashboard

This component generates counterfactual points for the supplied test dataset. It has a single input port, which accepts the output of the RAI Insights dashboard constructor. It also accepts the following parameters:

Parameter name Description Type
total_CFs The number of counterfactual points to generate for each row in the test dataset. Optional integer, defaults to 10.
method The dice-ml explainer to use. Optional string. Either random, genetic, or kdtree. Defaults to random.
desired_class Index identifying the desired counterfactual class. For binary classification, this should be set to opposite. Optional string or integer. Defaults to 0.
desired_range For regression problems, identify the desired range of outcomes. Optional list of two numbers3.
permitted_range Dictionary with feature names as keys and the permitted range in a list as values. Defaults to the range inferred from training data. Optional string or list3.
features_to_vary Either a string all or a list of feature names to vary. Optional string or list3.
feature_importance Flag to enable computation of feature importances by using dice-ml. Optional Boolean. Defaults to True.

3 For the non-scalar parameters: Parameters that are lists or dictionaries should be passed as single JSON-encoded strings.

This component has a single output port, which can be connected to one of the insight_[n] input ports of the Gather RAI Insights dashboard component.

 counterfactual_01: 
    type: command 
    component: azureml://registries/azureml/components/microsoft_azureml_rai_tabular_counterfactual/versions/<version>
    inputs: 
      rai_insights_dashboard: ${{parent.jobs.create_rai_job.outputs.rai_insights_dashboard}} 
      total_CFs: 10 
      desired_range: "[5, 10]"

Add Error Analysis to RAI Insights dashboard

This component generates an error analysis for the model. It has a single input port, which accepts the output of the RAI Insights Dashboard Constructor. It also accepts the following parameters:

Parameter name Description Type
max_depth The maximum depth of the error analysis tree. Optional integer. Defaults to 3.
num_leaves The maximum number of leaves in the error tree. Optional integer. Defaults to 31.
min_child_samples The minimum number of datapoints required to produce a leaf. Optional integer. Defaults to 20.
filter_features A list of one or two features to use for the matrix filter. Optional list, to be passed as a single JSON-encoded string.

This component has a single output port, which can be connected to one of the insight_[n] input ports of the Gather RAI Insights Dashboard component.

  error_analysis_01: 
    type: command 
    component: azureml://registries/azureml/components/microsoft_azureml_rai_tabular_erroranalysis/versions/<version>
    inputs: 
      rai_insights_dashboard: ${{parent.jobs.create_rai_job.outputs.rai_insights_dashboard}} 
      filter_features: `["style", "Employer"]'

Add Explanation to RAI Insights dashboard

This component generates an explanation for the model. It has a single input port, which accepts the output of the RAI Insights Dashboard Constructor. It accepts a single, optional comment string as a parameter.

This component has a single output port, which can be connected to one of the insight_[n] input ports of the Gather RAI Insights dashboard component.

  explain_01: 
    type: command 
    component: azureml://registries/azureml/components/microsoft_azureml_rai_tabular_explanation/versions/<version>
    inputs: 
      comment: My comment 
      rai_insights_dashboard: ${{parent.jobs.create_rai_job.outputs.rai_insights_dashboard}}

Gather RAI Insights dashboard

This component assembles the generated insights into a single Responsible AI dashboard. It has five input ports:

  • The constructor port that must be connected to the RAI Insights dashboard constructor component.
  • Four insight_[n] ports that can be connected to the output of the tool components. At least one of these ports must be connected.

There are two output ports:

  • The dashboard port contains the completed RAIInsights object.
  • The ux_json port contains the data required to display a minimal dashboard.
  gather_01: 
    type: command 
    component: azureml://registries/azureml/components/microsoft_azureml_rai_tabular_insight_gather/versions/<version>
    inputs: 
      constructor: ${{parent.jobs.create_rai_job.outputs.rai_insights_dashboard}} 
      insight_1: ${{parent.jobs.causal_01.outputs.causal}} 
      insight_2: ${{parent.jobs.counterfactual_01.outputs.counterfactual}} 
      insight_3: ${{parent.jobs.error_analysis_01.outputs.error_analysis}} 
      insight_4: ${{parent.jobs.explain_01.outputs.explanation}}

How to generate a Responsible AI scorecard (preview)

The configuration stage requires you to use your domain expertise around the problem to set your desired target values on model performance and fairness metrics.

Like other Responsible AI dashboard components configured in the YAML pipeline, you can add a component to generate the scorecard in the YAML pipeline:

scorecard_01: 

   type: command 
   component: azureml:rai_score_card@latest 
   inputs: 
     dashboard: ${{parent.jobs.gather_01.outputs.dashboard}} 
     pdf_generation_config: 
       type: uri_file 
       path: ./pdf_gen.json 
       mode: download 

     predefined_cohorts_json: 
       type: uri_file 
       path: ./cohorts.json 
       mode: download

Where pdf_gen.json is the score card generation configuration json file, and predifined_cohorts_json ID the prebuilt cohorts definition json file.

Here's a sample JSON file for cohorts definition and scorecard-generation configuration:

Cohorts definition:

[ 
  { 
    "name": "High Yoe", 
    "cohort_filter_list": [ 
      { 
        "method": "greater", 
        "arg": [ 
          5 
        ], 
        "column": "YOE" 
      } 
    ] 
  }, 
  { 
    "name": "Low Yoe", 
    "cohort_filter_list": [ 
      { 
        "method": "less", 
        "arg": [ 
          6.5 
        ], 
        "column": "YOE" 
      } 
    ] 
  } 
]

Here's a scorecard-generation configuration file as a regression example:

{ 
  "Model": { 
    "ModelName": "GPT-2 Access", 
    "ModelType": "Regression", 
    "ModelSummary": "This is a regression model to analyze how likely a programmer is given access to GPT-2" 
  }, 
  "Metrics": { 
    "mean_absolute_error": { 
      "threshold": "<=20" 
    }, 
    "mean_squared_error": {} 
  }, 
  "FeatureImportance": { 
    "top_n": 6 
  }, 
  "DataExplorer": { 
    "features": [ 
      "YOE", 
      "age" 
    ] 
  }, 
  "Fairness": {
    "metric": ["mean_squared_error"],
    "sensitive_features": ["YOUR SENSITIVE ATTRIBUTE"],
    "fairness_evaluation_kind": "difference OR ratio"
  },
  "Cohorts": [ 
    "High Yoe", 
    "Low Yoe" 
  ]  
}

Here's a scorecard-generation configuration file as a classification example:

{
  "Model": {
    "ModelName": "Housing Price Range Prediction",
    "ModelType": "Classification",
    "ModelSummary": "This model is a classifier that predicts whether the house will sell for more than the median price."
  },
  "Metrics" :{
    "accuracy_score": {
        "threshold": ">=0.85"
    },
  }
  "FeatureImportance": { 
    "top_n": 6 
  }, 
  "DataExplorer": { 
    "features": [ 
      "YearBuilt", 
      "OverallQual", 
      "GarageCars"
    ] 
  },
  "Fairness": {
    "metric": ["accuracy_score", "selection_rate"],
    "sensitive_features": ["YOUR SENSITIVE ATTRIBUTE"],
    "fairness_evaluation_kind": "difference OR ratio"
  }
}

Definition of inputs for the Responsible AI scorecard component

This section lists and defines the parameters that are required to configure the Responsible AI scorecard component.

Model

ModelName Name of model
ModelType Values in ['classification', 'regression'].
ModelSummary Enter text that summarizes what the model is for.

Note

For multi-class classification, you should first use the One-vs-Rest strategy to choose your reference class, and then split your multi-class classification model into a binary classification problem for your selected reference class versus the rest of the classes.

Metrics

Performance metric Definition Model type
accuracy_score The fraction of data points that are classified correctly. Classification
precision_score The fraction of data points that are classified correctly among those classified as 1. Classification
recall_score The fraction of data points that are classified correctly among those whose true label is 1. Alternative names: true positive rate, sensitivity. Classification
f1_score The F1 score is the harmonic mean of precision and recall. Classification
error_rate The proportion of instances that are misclassified over the whole set of instances. Classification
mean_absolute_error The average of absolute values of errors. More robust to outliers than mean_squared_error. Regression
mean_squared_error The average of squared errors. Regression
median_absolute_error The median of squared errors. Regression
r2_score The fraction of variance in the labels explained by the model. Regression

Threshold: The desired threshold for the selected metric. Allowed mathematical tokens are >, <, >=, and <=m, followed by a real number. For example, >= 0.75 means that the target for the selected metric is greater than or equal to 0.75.

Feature importance

top_n: The number of features to show, with a maximum of 10. Positive integers up to 10 are allowed.

Fairness

Metric Definition
metric The primary metric for evaluation fairness.
sensitive_features A list of feature names from the input dataset to be designated as sensitive features for the fairness report.
fairness_evaluation_kind Values in ['difference', 'ratio'].
threshold The desired target values of the fairness evaluation. Allowed mathematical tokens are >, <, >=, and <=, followed by a real number.
For example, metric="accuracy", fairness_evaluation_kind="difference".
<= 0.05 means that the target for the difference in accuracy is less than or equal to 0.05.

Note

Your choice of fairness_evaluation_kind (selecting 'difference' versus 'ratio') affects the scale of your target value. In your selection, be sure to choose a meaningful target value.

You can select from the following metrics, paired with fairness_evaluation_kind, to configure your fairness assessment component of the scorecard:

Metric fairness_evaluation_kind Definition Model type
accuracy_score difference The maximum difference in accuracy score between any two groups. Classification
accuracy_score ratio The minimum ratio in accuracy score between any two groups. Classification
precision_score difference The maximum difference in precision score between any two groups. Classification
precision_score ratio The maximum ratio in precision score between any two groups. Classification
recall_score difference The maximum difference in recall score between any two groups. Classification
recall_score ratio The maximum ratio in recall score between any two groups. Classification
f1_score difference The maximum difference in f1 score between any two groups. Classification
f1_score ratio The maximum ratio in f1 score between any two groups. Classification
error_rate difference The maximum difference in error rate between any two groups. Classification
error_rate ratio The maximum ratio in error rate between any two groups. Classification
Selection_rate difference The maximum difference in selection rate between any two groups. Classification
Selection_rate ratio The maximum ratio in selection rate between any two groups. Classification
mean_absolute_error difference The maximum difference in mean absolute error between any two groups. Regression
mean_absolute_error ratio The maximum ratio in mean absolute error between any two groups. Regression
mean_squared_error difference The maximum difference in mean squared error between any two groups. Regression
mean_squared_error ratio The maximum ratio in mean squared error between any two groups. Regression
median_absolute_error difference The maximum difference in median absolute error between any two groups. Regression
median_absolute_error ratio The maximum ratio in median absolute error between any two groups. Regression
r2_score difference The maximum difference in R2 score between any two groups. Regression
r2_Score ratio The maximum ratio in R2 score between any two groups. Regression

Input constraints

What model formats and flavors are supported?

The model must be in the MLflow directory with a sklearn flavor available. Additionally, the model needs to be loadable in the environment that's used by the Responsible AI components.

What data formats are supported?

The supplied datasets should be mltable with tabular data.

Next steps