RESEARCH · RESEARCH · #1119
Synthetic ground-truth framework for evaluating XAI methods (arXiv:2609.30397v1)
The paper proposes a framework that uses controlled interventions to generate synthetic datasets with known input-component importances, creating ground-truth explanations aligned with a model's behavior. The framework is instantiated on binary images, tabular data, and time series and is used to evaluate nine common XAI methods, revealing significant limitations in current techniques.
KEY POINTS
- The paper proposes a framework that uses controlled interventions to generate synthetic datasets with known input-component importances, creating ground-truth explanations aligned with a model's behavior.
- The framework is instantiated on binary images, tabular data, and time series and is used to evaluate nine common XAI methods, revealing significant limitations in current techniques.
- Provides an intervention-based way to construct ground-truth explanations, enabling more reliable assessment of XAI methods and exposing weaknesses that fidelity-based metrics can miss.
WHY IT MATTERS
Provides an intervention-based way to construct ground-truth explanations, enabling more reliable assessment of XAI methods and exposing weaknesses that fidelity-based metrics can miss.