Learning from Aggregate Responses

Adel Javanmard
University of Southern California (USC)

In many practical applications the training data is aggregated before being shared with the learner, in order to protect privacy of users' sensitive responses. In an aggregate learning framework, the dataset is grouped into bags of samples, where each bag is available only with an aggregate response, providing a summary of individuals’ responses in that bag. In this talk, I will discuss some of the recent developments in this space, namely on loss construction and bagging schemes which improve the accuracy of the model, while providing privacy. In particular, I will show how priors can be used to inform bag construction and also present an iterative boosting algorithm which refines the prior via sample splitting.

(This talk is based on collaboration with Google Research).


Back to EnCORE Workshop on Computational vs Statistical Gaps in Learning and Optimization