In my research on Trustworthy AI and intersectional fairness in computer vision, I've learned that model accuracy tells only half the story. When AI systems impact people's lives, opacity becomes an ethical liability.

The Tools of the Trade: SHAP & LIME in Practice

LIME (Local Interpretable Model-agnostic Explanations)

Creates a local surrogate model by perturbing inputs to approximate the black box’s behavior around a specific instance.
SHAP (SHapley Additive exPlanations)

Uses game theory (Shapley values) to assign each feature its fair contribution to a prediction. Offers strong theoretical guarantees.

Hands-On: Detecting Intersectional Bias

In my work on intersectional fairness in image classification, I used both tools to understand why models fail on certain subgroups.

Key Insight from SHAP:

Environmental features (lighting, background complexity) contributed 35% more to predictions for underrepresented intersections — a clear sign of bias.

Critical Analysis: Strengths and Limitations

Tool Strengths Limitations for Fairness
LIME Model-agnostic, intuitive, great for debugging specific cases Can be unstable; explanations vary across runs
SHAP Theoretically grounded, consistent, global + local views Computationally expensive; assumes feature independence

Key Finding from My Research

SHAP revealed that the baseline model relied 57% more on environmental features for underrepresented intersections. This led to the development of Bias-Weighted Augmentation.

Conclusion: From Explanations to Accountability

Technical explainability is necessary but not sufficient. We need to combine SHAP and LIME with rigorous fairness metrics, domain expertise, and engagement with affected communities to build truly equitable AI systems.