ContextualSHAP : Enhancing SHAP Explanations Through Contextual Language Generation

SHAP visualizations alone may not be sufficient for non-technical users: The paper identifies a key gap in existing explainability tools—while SHAP effectively shows feature importance through visualizations, these outputs often lack the contextual information that end-users without technical backgrounds need to understand AI decisions.

Natural language explanations can significantly improve comprehension: User evaluations in the healthcare case study found that when SHAP visualizations were paired with AI-generated textual explanations, participants perceived the explanations as "more understandable and contextually appropriate," suggesting that combining visual and narrative formats enhances clarity for diverse audiences.

Customization parameters make explanations more relevant: The proposed ContextualSHAP package allows users to define parameters like feature aliases, descriptions, and background context, enabling explanations to be tailored to specific domains and user perspectives rather than providing one-size-fits-all outputs.

LLM integration offers a practical path forward for XAI: By integrating large language models (specifically GPT) with established explainability methods like SHAP, the research demonstrates a feasible approach to making AI decisions more interpretable for stakeholders in high-stakes domains like healthcare.

Real-world user testing validates the approach: Rather than relying solely on technical metrics, the evaluation involved actual end-users and gathered feedback through Likert-scale surveys and interviews, providing evidence that the enhancement addresses genuine usability challenges in practice.