Data analysis
Surface imbalance across data and protected attributes, then correct it before training.
EXPLAIN→SURFACE BIAS→MITIGATE
CHECKED ✓OPEN SOURCE · ALPHA · MIT
XAI — an eXplainability toolbox for machine learning
"XAI is a Machine Learning library that is designed with AI explainability in its core." It implements the three steps of explainable machine learning: data analysis, model evaluation and production monitoring, tools and process together.
01 — THE PROCESS, BY CHECKPOINT
The process, by checkpoint
XAI converts the Responsible AI Principles into a practical tool + process approach. The process carries explainability and bias mitigation through the full responsible-ML workflow, rather than treating them as a one-off report after training.
TOOL + PROCESS
Responsible ML is a continuous workflow
Inspect the data before training
Model evaluation
Explain model behaviour and compare outcomes across protected groups.
EXPLAIN→SURFACE BIAS→MITIGATE
CHECKED ✓Production monitoring
Monitor predictions, surface bias and feed findings back into mitigation.
EXPLAIN→SURFACE BIAS→MITIGATE
CHECKED ✓RESPONSIBLE AI PRINCIPLESApplied at every checkpointPRODUCTION FEEDBACK ↺
02 — THE TOOLBOX, BY STAGE
The toolbox, by stage
The library ships with a census dataset for experimentation (xai.data.load_census()), and installs from PyPI. The protected-attribute lens runs through every stage: the same gender or ethnicity split that shapes the train/test split reappears in the metric and ROC comparisons.
STAGE 01 — PUBLIC API
Data analysis
imbalance_plot+
Plot class imbalances across one or more columns.
balance+
Correct class imbalances with up-sampling or down-sampling.
correlations+
Visualise correlations as a matrix or hierarchical dendogram.
group_by_columns+
Group a dataset by one or more categorical columns for custom operations.
balanced_train_test_split+
Create train and test splits balanced across protected attributes.
STAGE 02 — PUBLIC API
Model evaluation
feature_importance+
Calculate permutation-based feature importance.
metrics_plot+
Compare metric imbalances overall and across protected columns.
roc_plot+
Plot ROC curves overall or per group, such as gender, ethnicity, or age.
confusion_matrix_plot+
Render a confusion matrix for model predictions.
STAGE 03 — PUBLIC API
Production monitoring
smile_imbalance+
Bucket accuracy and statistical metrics by prediction probability to visualise benefits of adding manual review on probability thresholds.
03 — TALK AND WORKED EXAMPLE
Talk and worked example
The approach was presented at PyData London, covering how to identify and mitigate undesired bias with the toolbox.
PYDATA LONDON 2019 · VIDEOExplainable AI and bias mitigation with the XAI LibraryWatch the talk