Fairness
Last updated
Was this helpful?
Last updated
Was this helpful?
Humans naturally think in counterfactuals: "What if this had been different?" This thought process is particularly relevant in machine learning. For instance, if a customer churn prediction is high, a business might wonder what changes could retain the customer. What-if Analysis allows you to explore these scenarios effortlessly, answering the question: "How would the prediction have been if input X had been different?".
To begin with what-if analysis, open the View AI dashboard and navigate to the Fairness tab.
1. Select a Data Point
Start by choosing a row from the Data Points table. This will display the prediction for the selected data point, its probability, and a detailed breakdown of feature contributions, breaking down the model's decision-making process for a particular instance, highlighting the influence of each feature.
2. Change Input Features
Next, you can modify the input features to see how the model's prediction changes. Click Predict to see how these changes impact the prediction.
3. Toggle Counterfactual Analysis
Toggle the "Nearest Counterfactual" switch to find the minimal changes required for an opposite outcome (e.g., from churned to not churned).
Counterfactual-analysis helps you find the minimal changes needed to alter the prediction. For example, if a customer is predicted to churn, toggling the counterfactual analysis will show you what changes could make them stay.
4. Change Features for Counterfactuals
You can further refine your counterfactual-analysis by selecting which features to consider when generating counterfactuals. Click on the "Choose Features" dropdown to select or deselect features for generating counterfactuals. This is useful when some features are not feasible to change. For instance, if the counterfactual suggests changing "gender" to female, you can deselect "gender" to see other viable changes.
While performing what-if analysis, take advantage of the Copilot AI. This copilot can help you handle your data analysis tasks by providing insights, and even code snippets without needing to write complex queries or code.
When generating counterfactuals, deselect features you can't control to see how to achieve the opposite outcome with minimal changes. This approach helps identify actionable and practical adjustments.
