CUBE: A New Framework for Explainable AI Using Factorial Experimental Design
A new framework called CUBE (Contrastive Understanding by Balanced Experiments) has been developed by researchers to provide post-hoc explanations for black-box models using factorial experimental design. This framework assesses trained predictors through balanced combinations of low and high probes, summarizing the results as factorial effects. Main effects and pairwise interactions are viewed as controlled contrasts within a designated explanation region. Complete factorial probes accurately pinpoint these effects in the chosen design space, while fractional probes help lower query costs and reveal aliasing and resolution limitations. Experiments conducted on both synthetic and real tabular tasks show that CUBE effectively captures the primary learned effect structure and clarifies the limits of query-efficient explanations. This research is available on arXiv in the fields of computer science and machine learning.
Key facts
- CUBE stands for Contrastive Understanding by Balanced Experiments.
- It is a post-hoc explanation framework for black-box models.
- It uses factorial experimental design to generate explanations.
- Evaluates predictors on balanced low-high probe combinations.
- Summarizes responses as factorial effects.
- Main effects and pairwise interactions are interpreted as controlled contrasts.
- Complete factorial probes identify effects exactly on the design space.
- Fractional probes reduce query cost and expose aliasing constraints.
- Tested on synthetic and real tabular tasks.
- Published on arXiv under cs.LG.
Entities
Institutions
- arXiv