Enhancing Data Quality with Machine Learning: A Framework for Detecting and Revising Incorrect Labels

In this paper, the authors propose a new tool called pyAC to help data scientists create and manage Aspectual Concepts (ACs) for Machine Learning (ML) components. ACs are essential for ensuring that ML systems are safe, reliable, and compliant with regulations. However, creating and managing ACs can be complex and time-consuming, especially when working with different ML frameworks and systems.
The authors explain that traditional AC tools are mainly designed for safety engineers, but they may not be suitable for data scientists who work with Jupyter Notebooks and Python. To address this gap, the pyAC framework was developed to integrate with existing AC tools and provide a seamless user experience for data scientists.
The authors describe how the pyAC framework works by combining textual explanations, executable code blocks, and computational output in a Jupyter Notebook. This allows data scientists to create reproducible and understandable routines that can be easily shared with others. The framework also provides measures, which are lists of available blueprints that implement the AC. These blueprints contain justifications for why they can address the corresponding claim, as well as information on the model and data versions used.
The authors highlight three main contributions of their work: introducing the pyAC tooling framework, implementing AC elements in Jupyter Notebook and Python, and applying the framework to ensure the quality of test data. They also provide future directions for improving the framework and concludes the paper.
Throughout the article, the authors use everyday language and engaging metaphors to demystify complex concepts. For example, they explain that ACs are like recipes for ML systems, while pyAC is a tool that helps data scientists create and manage these recipes more efficiently. By using this approach, the authors make it easier for readers to understand the importance of ACs and how the pyAC framework can help them work more effectively with these concepts.