Skip to content

0.4.0 - 2024-01-29

📈️ Static and Interactive Visualizations

  • An interactive web app serves as a visualization component within the Virny model profiling library, empowering data scientists to engage in responsible model selection and to generate nutritional labels for their models. This application allows users to scrutinize dataset properties related to protected groups, compare models across diverse performance dimensions, and generate a comprehensive nutritional label for the most optimal model. The demonstration of the web app is hosted on the Hugging Face space. More details are in the Examples section of the documentation.

  • Improved and extended static visualizations provided by the Metric Visualizer.

⚙️ New Metric Computation Capabilities

  • A new capability to input an inprocessor into a metric computation interface as a basic model to profile an in-processing fairness intervention. Currently, only inprocessors from aif360 are supported. More details are in the Examples section of the documentation.

  • A new capability to input a postprocessor into a metric computation interface to use post-processing fairness interventions during model profiling. Currently, only postprocessors from aif360 are supported. More details are in the Examples section of the documentation.

💠 Analyzers and Metrics

  • Added a sample size for each protected group to an overall metrics matrix. Useful to know if the size of a protected group is big enough to be representative.

  • Simplified adding new metrics. Now, all functions, which compute overall metrics, are defined in Virny's metrics package.

  • Improved definition of disparity metrics. Now, all disparity metrics and their expressions are defined in the Metric Composer.

🗃 New Benchmark Fair-ML Dataset