Skip to content

Model Auditor

Understand where a binary classifier performs well, where errors concentrate, and how operating thresholds change the result.

Model Auditor is a Python library for evaluating fixed binary-classification predictions overall and across subgroups. It combines classification, ranking, probability, and decision metrics with explicit inference, confusion-group error analysis, paired comparisons, and visualization-ready outputs.

Use it when you already have one row per evaluated observation in a pandas DataFrame, including:

  • a binary ground-truth outcome;
  • one or more continuous score columns; and
  • categorical or discrete columns that define subgroups.
  • Compare sensitivity, specificity, AUROC, AUPRC, precision, F-scores, MCC, error rates, and support counts across feature levels.
  • Attach one scalar threshold to a score, override it for a single run, or apply different thresholds by feature level.
  • Select a scalar operating point with the Youden index or a minimum sensitivity or specificity target.
  • Configure Wilson, exact, IID, stratified, or cluster-resampled pointwise intervals with a local seed and diagnostics.
  • Compare paired models or feature levels with metric differences or ratios.
  • Evaluate probability calibration and descriptive decision net benefit.
  • Measure how strongly each subgroup is represented in true-positive, true-negative, false-positive, and false-negative sets.
  • Export unrounded long-form results with support and provenance, display tables, or neutral-by-default pandas Styler objects.
  • Plot metric intervals and subgroup score distributions with Matplotlib.
  • Generate node arrays for sunburst or treemap charts rendered by Plotly or another library.
  • Define custom metrics over the package’s recognized truth, score, and confusion-input columns.
Your goal Start here
Install the package and run one verified evaluation Getting started
Compare model performance across subgroups Evaluate subgroup performance
Compare two scores or feature levels Compare models and groups
Check probability calibration or net benefit Evaluate calibration and decisions
Choose an operating threshold Optimize a scalar threshold
Apply policy-specific thresholds by region, cohort, or channel Apply conditional thresholds
Find subgroups concentrated in false positives or false negatives Analyze confusion-group representation
Prepare notebook tables or publication inputs Export and style results
Understand exact method signatures and defaults Reference
Work on the package or synchronize this site Contributing

Model Auditor v0.1.16 evaluates binary outcomes. It does not train models, choose protected attributes for you, determine whether a disparity is acceptable, or convert relative table colors into a fairness conclusion. Threshold optimization uses the data you provide, so operating-point selection and final evaluation should be separated when unbiased performance estimates matter.

See package status and compatibility and known edge cases before adopting the package in a production workflow.