0.1.0 - 2023-02-06¶
🚀 Models Audit Pipeline¶
-
Developed an entire pipeline for auditing model stability and fairness with detailed reports and visualizations
-
Designed and implemented an extensible architecture split on components (User interfaces, MetricsComposer, etc.) that can be easily adapted to your needs
-
Enabled easy pipeline adaptability for different classification datasets
-
Added a feature to audit blind classifiers, which were trained on features without sensitive attributes, and use these sensitive attributes for analysis
👩💻 User Interfaces¶
-
Added three types of user interfaces:
- Interface for multiple runs and multiple models
- Interface for multiple models and one run
- Interface for one model and one run
-
Added an ability to input arguments to interfaces via user-friendly config yaml files or direct arguments
🗃 Datasets and Preprocessing¶
-
Added built-in preprocessing techniques for raw classification datasets
-
Developed an ability to work with non-binary features
-
Enabled access to COMPAS and Folktables benchmark datasets via implemented data loaders
💠 Analyzers and Metrics¶
-
Added an ability to analyze intersections of sensitive attributes
-
Implemented a set of error and variance metrics:
- 6 subgroup variance metrics
- Mean
- Std
- IQR
- Entropy
- Jitter
- Label Stability
- 8 subgroup error metrics
- TPR
- TNR
- PPV
- FNR
- FPR
- Accuracy
- F1
- Selection-Rate
- 5 group variance metrics
- Label Stability Ratio
- IQR Difference
- Std Difference
- Std Ratio
- Jitter Difference
- 5 group fairness metrics
- Equalized Odds TPR
- Equalized Odds FPR
- Disparate Impact
- Statistical Parity Difference
- Accuracy Difference
- 6 subgroup variance metrics
📈 Reports and Visualizations¶
-
Added an ability to create predefined plots for result metrics
-
Developed a feature to make detailed summary reports with visualizations
😌 Convenience¶
-
Enabled smart saving of result metrics in files
-
In the multiple runs interface, a file with result metrics is saved each time when each run is completed. In such a way, if you get an error in one of the runs, the results of the previous runs will be saved.
-
Enabled library installation via pip
-
Created and hosted a website for detailed documentation with examples