Easy Learning from Label Proportions

Robert Busa-Fekete
Travis Dick
Heejin Choi


We present a novel way of training models in theweakly supervised setup of learning frombagsof examples with just aggregate label informa-tion. Unlike previous work, we focus on learninga classifier that can produce accurate predictionsat an individual instance level as opposed to simply matching a bag’s label proportion. The mainstrength of our algorithm lies on the simplicity ofits implementation. We demonstrate that a simplemodification to the loss function can yield accu-rate event level estimates. Moreover we show thatthe sample complexity of doing empirical riskminimization or stochastic gradient descent withlabel proportions increases only by a factor of √k compared to traditional supervised learning sce-narios. Finally, we validate our theoretical resultson multiple datasets and demonstrate that our pro-posed algorithm beats provides state of the artperformance for learning with label proportions.

Research Areas