TLDR: ProfileXAI is a novel framework that combines traditional AI explanation methods (SHAP, LIME, Anchor) with retrieval-augmented Large Language Models (LLMs) to create personalized and easy-to-understand explanations for different user types, including machine learning engineers, domain experts, and non-technical users. It dynamically selects the most suitable explanation method for each instance and generates grounded narratives, leading to stable token usage and high user satisfaction, thereby enhancing AI transparency and trustworthiness.
In the rapidly evolving landscape of artificial intelligence, models are becoming increasingly complex, often operating as ‘black boxes’ where their decision-making processes are opaque. This lack of transparency poses a significant challenge, especially when AI systems are deployed in critical domains like healthcare. Explainable AI (XAI) aims to address this by making AI decisions understandable, but existing methods often fall short in adapting explanations to diverse audiences with varying levels of expertise.
Introducing ProfileXAI: User-Adaptive Explainable AI
A groundbreaking framework called ProfileXAI has emerged to tackle this challenge head-on. Developed by Gilber A. Corrales, Carlos Andrés Ferro Sánchez, Reinel Tabares-Soto, Jesús Alfonso López Sotelo, Gonzalo A. Ruz, and Johan Sebastian Piña Durán, ProfileXAI is designed to provide user-adaptive explanations for AI models. It’s a versatile system, independent of specific models or domains, that combines established post-hoc explainers like SHAP, LIME, and Anchor with advanced retrieval-augmented Large Language Models (LLMs).
The core idea behind ProfileXAI is to tailor explanations to different user profiles: machine learning engineers, domain experts, and non-technical users. This ensures that the explanations are not only accurate but also relevant and comprehensible to the intended audience, maximizing their utility and fostering trust in AI systems.
How ProfileXAI Works
The ProfileXAI architecture is a sophisticated pipeline that begins with a multimodal knowledge base, which supplies contextual information to enrich explanations. This knowledge base is processed, and relevant components (text, images) are stored in a vector database. When an AI model’s decision needs explaining, the system retrieves pertinent information from this database to compose a generation prompt for the LLM, enabling it to produce natural language explanations with rich context.
Crucially, ProfileXAI doesn’t rely on a single explanation method. Instead, an intelligent Explanation Engine dynamically selects the most suitable XAI method (SHAP, LIME, or Anchor) for each specific instance based on predefined quantitative criteria. This dynamic selection ensures that the explanation generated is the most appropriate for the given context and data.
Once an explanation is generated, it is then customized for one of three user profiles:
-
ML Engineer: Receives technical details, performance metrics, and raw model outputs.
-
Domain Expert: Gets explanations translated into terminology aligned with their specific application domain.
-
Non-technical User: Provided with accessible language, illustrative examples, and minimal jargon.
Furthermore, an interactive chat module allows users to ask follow-up questions, enabling a deeper exploration of any aspect of the generated explanation and resolving residual uncertainties in real-time.
Evaluating the Framework
The researchers rigorously evaluated ProfileXAI using two public datasets: Heart Disease (13 features) and Differentiated Thyroid Cancer Recurrence (16 features). They trained a multilayer perceptron (MLP) on the first dataset and a Random Forest on the second. The evaluation focused on three key areas:
-
XAI-metric analysis: Assessing fidelity, robustness, and effective complexity.
-
Token consumption: Measuring the computational cost and brevity of explanations across user profiles.
-
Satisfaction simulation: Using a simulated LLM to score explanation quality based on the Hoffman survey, a standard for evaluating human-AI interaction.
Key Findings and Insights
The study revealed that no single explainer method (SHAP, LIME, Anchor) universally dominates across all evaluation criteria. Each method demonstrated unique strengths:
-
LIME achieved the best balance between fidelity (how well the explanation reflects the model’s behavior) and robustness (stability of the explanation), though at a moderate complexity.
-
Anchor was found to produce the most parsimonious explanations, requiring the fewest features to alter a prediction, and generally resulted in the lowest token consumption, making it ideal when brevity is paramount for technical users.
-
SHAP attained high-fidelity explanations and consistently received the highest user satisfaction ratings across both datasets and user profiles, despite having diminished robustness and greater explanatory complexity.
A significant finding was that ProfileXAI’s profile conditioning mechanism successfully stabilized token usage across explanations (with standard deviations of token counts remaining low) and maintained consistently positive user satisfaction ratings (averaging ≥3.7 on a 1-5 scale) across all user profiles, including domain experts who were the most critical. This indicates that tailoring explanations to the user effectively manages cognitive load and enhances perceived quality.
Also Read:
- Enhancing Trust in Multimodal AI Through Consistent Emotional Explanations
- EGO-Prompt: Automating LLM Adaptation for Specialized Tasks with Evolving Domain Knowledge
Conclusion
ProfileXAI represents a significant advancement in Explainable AI by offering a dynamic, user-adaptive framework. By intelligently selecting XAI methods and verbalizing their outputs at an appropriate technical depth for different user profiles, the system successfully reconciles interpretability, cognitive economy, and stakeholder satisfaction. This approach ensures that AI explanations are not just accurate, but also truly useful and understandable to everyone, from machine learning specialists to non-technical stakeholders. For more details, you can refer to the original research paper.
