Model Explainability

Analyze the importance of each input field to your model's predictions using model explainability. Create a Workflow Classification model that includes a graphical analysis of feature importance by executing the provided script.

Avant de commencer

• This method uses the Workflow Classification Solution API, instead of the Solution Definition form, to create and train a model with explainability added. For information about the components of Workflow Classification models, see Create and train a classification solution.
• Role required: ml_admin or admin

Pourquoi et quand exécuter cette tâche

Model explainability helps identify the key features that influence the model's predictions during training.

Remarque :

Explainability can't be added to an existing model. This method uses a script to create and train a new Workflow Classification model. For more information about scripting the creation of Classification solutions, see ClassificationSolution - Global.

The script provided in the procedure creates and trains a model with explainability set to true. On the new model's solution form, an additional tab labeled Feature Importance appears. This tab offers a graph of the relative contribution of each input to the prediction.

Procédure

1. Navigate to All > System Definition > Scripts - Background.
2. Edit the query filter and table, field, and variable values in the following script according to your planned model, then execute the script.

   // Define a dataset
                       var myIncidentData = new sn_ml.DatasetDefinition({
                       'tableName': 'incident',
                       'fieldNames': ['category', 'short_description', 'sys_updated_by', 'assignment_group', 'description', 'priority'],
                       'encodedQuery': 'activeANYTHING'
                       });
                       
                       // Define a classification solution definition with explainability field
                       var mySolution = new sn_ml.ClassificationSolution({
                       'label': 'model explainability',
                       'dataset': myIncidentData,
                       'predictedFieldName': 'category',
                       'inputFieldNames': ['short_description', 'priority'],
                       //setting the explainability field to true.
                       'explainability': true,
                       });
                       
                       // Add solution to ClassificationSolutionStore
                       var my_unique_name = sn_ml.ClassificationSolutionStore.add(mySolution)
                       
                       // Submit training job
                       var solutionVersion = mySolution.submitTrainingJob();
                   

   Remarque :

   Substitute the query filter and table, field, and variable names in this script with your own values.
3. Navigate to the ML Solutions [ml_solution] table and open your new solution by selecting the value of its Active field.
4. On the solution form, locate and open the Feature Importance tab.
   Feature Importance displays a drop-down list.

   

   • The label for this drop-down list is the name of your output (predicted) field, so the label is specific to each model.
   • The values in the drop-down list are the possible output classes for your output field, plus the Global option.

   List option	Description
   Global	Provides an overview of how the model behaves across all predictions, showing the overall impact of each input feature. Select Global to open a graph of the importance of your input fields to predictions for all output classes as a whole.
   Your output class value	Focuses on the model's behavior for the chosen class only, showing how input features contribute to predictions on a per-class basis. Select one of the possible output classes to open a graph of the importance of your input fields to predictions for that class.

5. Launch the graphical analysis by selecting a value from the drop-down list.
   The y axis shows your input fields and the x axis shows the numerical importance for each input. The label of the graph reflects the class that you chose in the drop-down list.
   You can hover over a bar to display the numerical percentage for each input's importance.

Résultats

A positive importance value means that the input field increases the model's prediction score. A negative value means that the input field decreases the prediction score.

Que faire ensuite

Consider dropping input fields with low importance scores. Retrain your model after modification.

Concepts associés

• Using Machine Learning APIs

Tâches associées

• Create and train a classification solution

Information associée

• ClassificationSolution - Global