Ultra-marginal Feature Importance: Learning from Data with Causal Guarantees

UMFI provides causal insights, not just predictive accuracy: Unlike traditional machine learning approaches that focus on building the best-performing models, this work prioritizes understanding the actual causal relationships in data. The paper proves that UMFI satisfies specific axioms for explaining both causal and associative relationships, offering more trustworthy explanations of why features matter rather than just that they matter.

Better performance when real-world data is messy: UMFI outperforms the previous MCI method "especially in the presence of correlated interactions and unrelated features"—conditions that are common in real datasets. This means the approach is more robust to the kinds of data complexity that often trip up simpler feature importance methods.

Significant speed improvements make it practical: By reducing runtime from exponential to super-linear, UMFI becomes computationally feasible for larger datasets. This matters for responsible AI because impractical methods often get abandoned in favor of simpler but less reliable alternatives.

Bridges fairness and interpretability research: The method draws on dependence removal techniques from AI fairness literature, suggesting that tools developed to reduce bias can also help us better understand causal relationships in data—a useful cross-pollination between different areas of responsible AI.

Moves ML research toward scientific understanding: The paper explicitly addresses a gap where machine learning prioritizes predictive performance over learning what data actually reveals, offering a framework that serves scientists and decision-makers who need to understand relationships, not just make predictions.