VTruST: Controllable value function based subset selection for Data-Centric Trustworthy AI

Data selection directly impacts AI trustworthiness: The paper demonstrates that carefully choosing which training data to use can improve multiple trustworthiness metrics simultaneously—fairness, robustness, and accuracy—rather than requiring separate fixes after model development.

Users can control trade-offs between competing goals: VTruST's framework allows practitioners to adjust priorities when trustworthiness metrics conflict with each other, giving stakeholders agency in deciding what matters most for their specific application (e.g., prioritizing fairness over raw accuracy in high-stakes decisions).

The approach works across diverse domains: Experimental validation on "social, image, and scientific datasets" suggests this data-centric approach is broadly applicable rather than limited to one type of AI application, making it potentially useful for various high-stakes fields.

Data values provide transparency: The framework generates explanations showing why certain data points matter for different trustworthiness metrics, addressing the "black box" problem by making data selection decisions interpretable to users and stakeholders.

Efficiency matters for practical adoption: By developing an efficient online algorithm for subset selection, the paper addresses a real implementation challenge—making trustworthy AI practical means the solution must work with reasonable computational resources, not just in theory.