KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Global Feature Importance
Aggregated metric evaluating the impact of each input variable on the model's predictions across the entire dataset, essential for understanding general fraud risk factors.
Local Feature Importance
Measure of each variable's influence on a specific prediction, crucial for investigating individual transactions flagged as fraudulent by the model.
Extracted Association Rules
Set of logical conditions (e.g., IF amount > X AND country = Y THEN high risk) extracted from a 'black box' model to provide understandable and actionable explanations for anti-fraud analysts.
Surrogate Decision Tree
Simple model (decision tree) trained to mimic the predictions of a complex model, serving as an interpretable proxy to globally explain the logic of the anti-fraud model.
Anchor Explanations
High-precision decision rules that 'anchor' a prediction, explaining why a model made a specific decision for a transaction by identifying sufficient conditions.
Explanation Fidelity
Metric evaluating how well a local explanation (e.g., LIME) faithfully represents the behavior of the complex model in the vicinity of a specific prediction.
Dependence Map
Two-dimensional visualization showing how the fraud prediction changes based on the interaction between two features, revealing complex risk synergies.
Automated Justification Report
Document generated by the XAI system consolidating explanations (SHAP values, anchors, counterfactuals) for a flagged transaction, intended for investigators and regulatory compliance.
Feature Sensitivity Analysis
Process evaluating how minor variations in input features affect the fraud risk score, essential for testing the model's robustness against manipulation.