Reproduce “Optimal Learning from Label Proportions with General Loss Functions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lorne Applebaum, Travis Dick, Claudio Gentile, Haim Kaplan, Tomer Koren
Researchers from Google Research and Tel Aviv University developed a debiasing methodology for Learning from Label Proportions (LLP) that accommodates general loss functions, including unbounded ones. The method achieves sample complexity guarantees where estimator variance is independent of bag size, demonstrating improved performance for large aggregate data in both batch and online experiments.
Key insights
- The core debiasing technique provides a bag-level loss estimator with variance independent of the bag size (k), marking a critical advancement over prior methods where variance scaled linearly with k.
- Achieved improved sample complexity rates: a 'slow rate' of O(k/β²) for non-realizable settings and an optimal 'fast rate' of O(k/β) for realizable settings with quadratically sandwiched losses.
- Provided the first unified theoretical treatment for multi-class LLP, extending the methodology to scenarios with varying granularities of aggregate class information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.