Graph VAEs Improve Noisy Crowdsourced Label Aggregation

This paper introduces a novel approach for aggregating noisy labels from crowdsourced data, a common challenge in machine learning. Unlike traditional Bayesian methods that rely on predefined generative models or neural network techniques requiring pseudo-labels, this new method, called Dual-Primal Graph VAEs, offers an unsupervised solution. The architecture leverages Graph Attention Networks (GATs) for message passing on both the primary graph of the crowdsourced dataset and its dual. By treating the true labels as latent variables, the model can learn representations without needing a separate classifier or synthetic data. Experimental results demonstrate that this model achieves state-of-the-art performance on various crowdsourcing benchmarks. Furthermore, the framework is flexible enough to incorporate additional side information, such as representations from classifiers trained on the noisy labels, to significantly enhance classification performance during testing. This makes it a powerful tool for improving data quality in real-world applications.