Learning from label proportions (LLP) is a weakly supervised setting where training data are grouped into bags, and only the proportion of labels within each bag is observed. The goal is still to learn a classifier that predicts labels for individual instances, but without access to per-instance ground truth.
A new arXiv preprint, "Rethinking Learning from Label Proportions via Moment Matching," proposes a fresh perspective on this problem. The abstract indicates that the authors frame LLP as a moment-matching task, which may offer a more principled way to recover the underlying label distribution from bag-level statistics.
Because the abstract is truncated, specific methodological details and experimental results are not available from the source. The paper appears to be a theoretical contribution that rethinks the LLP formulation, but readers should consult the full manuscript for concrete claims and evidence.