Towards a statistical theory of data selection under weak supervision

TL;DR

This paper develops a statistical framework for selecting informative data subsets under weak supervision, demonstrating that strategic data selection can outperform using the full dataset and highlighting limitations of existing methods.

Contribution

It introduces a theoretical approach to data selection under weak supervision, combining mathematical analysis with experiments to improve understanding of optimal sampling strategies.

Findings

Data selection can outperform full dataset training in some cases.

Popular data selection methods may be suboptimal.

Theoretical insights guide better data sampling strategies.

Abstract

Given a sample of size $N$, it is often useful to select a subsample of smaller size $n<N$ to be used for statistical estimation or learning. Such a data selection step is useful to reduce the requirements of data labeling and the computational complexity of learning. We assume to be given $N$ unlabeled samples $\{{\boldsymbol x}_i\}_{i\le N}$, and to be given access to a `surrogate model' that can predict labels $y_i$ better than random guessing. Our goal is to select a subset of the samples, to be denoted by $\{{\boldsymbol x}_i\}_{i\in G}$, of size $|G|=n<N$. We then acquire labels for this set and we use them to train a model via regularized empirical risk minimization. By using a mixture of numerical experiments on real and synthetic data, and mathematical derivations under low- and high- dimensional asymptotics, we show that: $(i)$~Data selection can be very effective, in particular beating training on the full sample in some cases; $(ii)$~Certain popular choices in data selection methods (e.g. unbiased reweighted subsampling, or influence function-based subsampling) can be substantially suboptimal.