Entropic Projection Alignment: New Method for Distribution Shift
A new framework called Entropic Projection Alignment (EPA) addresses three key challenges of distribution shift: estimating model performance on unlabeled target domains, explaining the shift by identifying responsible features, and improving target domain performance. EPA aligns source distribution to target by matching selected moments while minimizing KL divergence from the source, yielding a unique closed-form solution for importance weights with implicit variance control. Drawing on domain adaptation theory, the method establishes that moment matching suffices for reliable estimation and adaptation, avoiding full density ratio recovery. Extensive experiments and theoretical guarantees show EPA consistently outperforms state-of-the-art baselines.
Key facts
- EPA addresses three challenges: performance estimation, shift explanation, and performance improvement under distribution shift.
- EPA aligns source to target by matching selected moments and minimizing KL divergence.
- The method yields a unique closed-form solution for importance weights.
- EPA achieves robustness through implicit variance control.
- Moment matching is sufficient for reliable estimation and adaptation.
- EPA avoids the need for full density ratio recovery.
- Extensive experiments show EPA outperforms state-of-the-art baselines.
- The framework is supported by strong theoretical guarantees.
Entities
—