Using Stratified Sampling to Improve LIME Image Explanations

LIME Image explanations can be unreliable due to sampling problems: The paper identifies that the current method for generating explanations uses random sampling that often misses important variations in the data, leading to explanations that may not accurately reflect how an AI model actually makes decisions about images.

Stratified sampling offers a practical fix: By using a more systematic sampling approach (stratified sampling) instead of random sampling, the researchers demonstrate that LIME Image can produce more consistent and trustworthy explanations without requiring changes to the underlying AI models being explained.

The problem affects real-world AI transparency: These sampling artifacts mean that when organizations use LIME Image to explain their computer vision systems to stakeholders or regulators, they may be presenting explanations that don't fully capture how the model works, potentially masking biases or errors.

Connection to broader explainability theory: The paper links this technical fix to Shapley theory, a foundational concept in responsible AI that emphasizes the importance of fairly representing all relevant factors—suggesting this sampling issue is a known concern across different explanation methods.

Improved explanations support better AI governance: More accurate explanations enable better auditing, debugging, and accountability of AI systems, which are essential for responsible deployment in high-stakes applications like healthcare, hiring, or criminal justice.