TL;DR
Optimal Transport (OT) label transfer often fails to identify rare classes, even when they are perfectly separated in embedding space. This paper reveals that the culprit is the "transport capacity" of the source marginal—essentially, rare classes simply run out of mass to give—and shows that matching this marginal to the target composition can double transfer performance.
What Problem Does This Paper Address?
Optimal Transport is a powerhouse for aligning distributions, especially in fields like single-cell genomics and image domain adaptation. However, it has a notorious "rare class" problem: if a class is infrequent in your source data but common in your target data (a "composition shift"), OT systematically under-predicts it.
Previously, researchers often blamed the quality of the data representation (the embedding). This paper argues that the failure is structural. In standard OT, every source point is weighted equally, meaning a class’s ability to "claim" target points is strictly limited by its frequency in the source. When target demand exceeds this "marginal capacity," the OT solver is forced to misallocate mass, leading to confident but incorrect predictions.
Methods
The authors investigate Entropic Unbalanced Optimal Transport (UOT). The key innovation is moving away from the standard uniform source marginal ($\mu$).
- Identifying the Failure: They use partial Spearman correlations to show that displaced mass from "starved" classes flows toward high-capacity (abundant) classes rather than geometrically similar ones.
- Marginal Correction: They propose reshaping the source marginal to reflect the target composition ($\pi$). Instead of uniform weights, source points are weighted such that the total mass available for class $k$ matches its expected frequency in the target.
- Estimation: Since target labels are unknown in practice, they test several "deployable" estimators for the target composition, including label-shift estimators (BBSE, RLLS) and transport-native ones like JCPOT and a simple kNN pre-pass.
- Double Dissociation: They develop a framework to decompose transfer error into a removable component (driven by capacity/coverage) and a residual component (driven by geometric separability).
Datasets
The study evaluates these concepts across two distinct domains:
- Single-cell Cross-modal Transfer: Aligning dissociated RNA sequencing (source) with spatial transcriptomics (target). They use the Jint (in-house) and Janesick (public) datasets.
- Image Domain Adaptation: Using the Office-Home benchmark, where they artificially vary class coverage to strictly isolate the effects of composition shift from stylistic shifts.
- SpatialCorpus: Used as a negative control where no composition shift exists.
Key Insights and Ideas
- Capacity is not Geometry: A class can be perfectly isolated in an embedding, but if it lacks "transport mass," the OT coupling will ignore it. This proves that transport capacity and representation geometry are distinct, independent sources of error.
- The "Upstream" Bottleneck: The authors show that you cannot fix this problem by tweaking the "read-out" (how you interpret the transport plan) or by adding more entropic smoothing. The failure happens in the coupling itself. Once the mass is misallocated by the solver to satisfy marginal constraints, the information is lost.
- Confident Hallucinations: Starved classes don't just result in "low confidence" predictions. The displaced mass lands confidently inside large, well-supported classes, making these errors invisible to standard uncertainty detectors.
- The Power of 30%: You don't need a perfect estimate of the target composition to see gains. Moving the marginal just 30% of the way toward the true target composition already recovers half of the potential performance gain.
Key Results
The impact of marginal correction is dramatic, particularly for rare classes:
- Performance Jump: On single-cell data, using an "oracle" (true) target composition raised the macro-F1 score from 0.322 to 0.623.
- Closing the Gap: The best practical estimator (kNN pre-pass) closed 69% of the gap between the baseline and the oracle on single-cell data. On Office-Home, the JCPOT estimator closed 59% of the gap.
- Error Separation: In the Office-Home experiments, the authors successfully "dissociated" the errors. They showed that their correction specifically removed the error associated with class coverage ($ \rho = -0.809 $) while leaving the error associated with how well classes are separated in space ($ \rho = -0.337 $).
Limitations and Future Work
- Estimator Dependency: The method's practical success relies heavily on the accuracy of the target composition estimate. In measurement-heavy shifts (like single-cell), standard label-shift estimators struggled, suggesting a need for more robust composition estimators.
- Selective Application: The correction is a double-edged sword. While it helps "starved" classes, it can slightly degrade performance for classes that already have sufficient coverage. The authors suggest the correction should be applied selectively.
- Fixed Representations: This study treated the data embeddings as fixed. A promising future direction is integrating marginal correction into the training of the encoders themselves to improve cross-domain alignment dynamically.
Summary updated