What-If Tool

What-If Tool analyzes machine learning model behavior to interpret predictions, assess feature effects, and measure fairness across hypothetical and real input scenarios.

Key Features:

• Interactive Probing: Explores how systematic changes to input data affect model predictions and outputs.
• Hypothetical Scenario Testing: Evaluates model responses to hypothetical or counterfactual input scenarios to probe potential behaviors and limitations.
• Feature Importance Analysis: Computes and presents importance of input features to quantify their influence on model predictions.
• Model Behavior Visualization: Visualizes model outputs across multiple models and subsets of input data for comparative analysis.
• Fairness Metrics Measurement: Calculates and reports various fairness metrics to assess equity across demographic or subgroup partitions.

Scientific Applications:

• Bias Identification: Identifies algorithmic biases by comparing performance and outcomes across data subgroups.
• Fairness Assessment: Assesses and quantifies fairness of decision-making using group- or subgroup-level metrics.
• Robustness and Sensitivity Analysis: Evaluates model robustness and sensitivity by testing predictions under varied and hypothetical input conditions.

Methodology:

Performs feature importance analysis, computes fairness metrics, tests models on hypothetical or counterfactual inputs, and visualizes outputs across models and data subsets.