TS-Insight: Visualizing Thompson Sampling for Verification and XAI

Thompson Sampling algorithms are difficult to understand by default: The paper identifies that Thompson Sampling and its variants, despite being widely used in active learning systems, operate as "black boxes" due to their probabilistic nature, making it hard for developers to debug problems or understand why specific decisions were made.

Visualization can make AI decision-making more transparent: TS-Insight demonstrates that visual analytics tools can effectively illuminate the internal mechanisms of complex algorithms by tracking how each option's probability estimates evolve over time, what evidence supports those estimates, and how sampling outcomes influence decisions—directly addressing the "explainability" requirement for responsible AI.

Transparency tools are especially critical for high-stakes applications: The paper emphasizes that interpretable decision-making is particularly important in "sensitive domains," suggesting that systems using Thompson Sampling in healthcare, criminal justice, hiring, or other consequential areas need these kinds of verification tools to ensure decisions can be audited and justified.

Developer-focused tools are necessary for responsible deployment: By explicitly designing TS-Insight "for model developers," the paper highlights that responsible AI requires not just end-user explanations, but also internal tools that enable technical teams to verify algorithms work as intended, diagnose failures, and build appropriate trust before deployment.