A Framework for Deprecating Datasets: Standardizing Documentation, Identification, and Communication
This addresses a critical gap in data stewardship for the ML community, though it is incremental as it builds on existing documentation practices.
The paper tackles the lack of standardized practices for deprecating datasets in machine learning, proposing a framework and centralized repository to address technical, legal, ethical, and organizational issues, with examples of datasets that continued circulating after deprecation.
Datasets are central to training machine learning (ML) models. The ML community has recently made significant improvements to data stewardship and documentation practices across the model development life cycle. However, the act of deprecating, or deleting, datasets has been largely overlooked, and there are currently no standardized approaches for structuring this stage of the dataset life cycle. In this paper, we study the practice of dataset deprecation in ML, identify several cases of datasets that continued to circulate despite having been deprecated, and describe the different technical, legal, ethical, and organizational issues raised by such continuations. We then propose a Dataset Deprecation Framework that includes considerations of risk, mitigation of impact, appeal mechanisms, timeline, post-deprecation protocols, and publication checks that can be adapted and implemented by the ML community. Finally, we propose creating a centralized, sustainable repository system for archiving datasets, tracking dataset modifications or deprecations, and facilitating practices of care and stewardship that can be integrated into research and publication processes.