crowd-hpo: Realistic Hyperparameter Optimization and Benchmarking for Learning from Crowds with Noisy Labels

Crowdworking is a cost-efficient solution to acquire class labels. Since these labels are subject to noise, various approaches to learning from crowds have been proposed. Typically, these approaches are evaluated with default hyperparameters, resulting in suboptimal performance, or with hyperparameters tuned using a validation set with ground truth class labels, representing an often unrealistic scenario. Moreover, both experimental setups can produce different rankings of approaches, complicating comparisons between studies. Therefore, we introduce crowd-hpo as a realistic benchmark and experimentation protocol including hyperparameter optimization under noisy crowd-labeled data. At its core, crowd-hpo investigates model selection criteria to identify well-performing hyperparameter configurations only with access to noisy crowd-labeled validation data. Extensive experimental evaluations with neural networks show that these criteria are effective for optimizing hyperparameters in learning from crowds approaches. Accordingly, incorporating such criteria into experimentation protocols is essential for enabling more realistic and fair benchmarking.