xai

Why we included this project

Teams shipping machine learning models into regulated settings like lending or healthcare need a practical way to spot where a model performs differently across protected groups. This library bundles the essentials into one importable package: it flags class imbalances on protected columns, rebalances data with upsampling or downsampling, computes permutation feature-importance scores, and draws ROC, precision-recall, and metric plots broken down by demographic groups. The probability-threshold tooling is useful when you want to reason about where a manual-review step adds value instead of blindly trusting a single cutoff. It follows the three-stage loop of data analysis, model evaluation, and production monitoring, so it drops into an existing workflow without a heavyweight platform. Data scientists and ML engineers checking bias or explainability early in a project will find the quickstart examples and bundled census dataset make it easy to try these patterns on their own data.