The Label Complexity of Class-Conditional Coverage under Distribution Shift
Per-class coverage under shift cannot be fixed for free; the paper pins down how many target labels it takes.
Han and Qu show that split conformal prediction can preserve overall coverage while quietly missing badly on individual classes. On a cross-subject skeleton benchmark, marginal coverage stays near 90%, but the worst action class is covered only about 70% and ten of sixty classes fall below 80%. They prove that with joint covariate-and-label shift, source labels plus unlabeled target data do not identify the target class-conditional score law. The label cost for both validity and efficiency scales with inverse squared tolerance and the log of the number of classes, while pseudo-labels offer only a small constant-factor efficiency gain where coverage collapses. ArXiv · AI/CL/LG's note
Han and Qu show that split conformal prediction can preserve overall coverage while quietly missing badly on individual classes. On a cross-subject skeleton benchmark, marginal coverage stays near 90%, but the worst action class is covered only about 70% and ten of sixty classes fall below 80%. They prove that with joint covariate-and-label shift, source labels plus unlabeled target data do not identify the target class-conditional score law. The label cost for both validity and efficiency scales with inverse squared tolerance and the log of the number of classes, while pseudo-labels offer only a small constant-factor efficiency gain where coverage collapses. ArXiv · AI/CL/LG's note
score 4