name: mitigation-algorithms
description: "Choose and run AIF360 legacy bias mitigation algorithms with
correct lifecycle stage, data contracts, optional extras, and metric checks."
disable-model-invocation: true
metadata:
disco-role: operating
license: Apache 2.0
AIF360 Mitigation Algorithms
Use this sub-skill when a task asks for legacy aif360.algorithms mitigation with AIF360 datasets: preprocessing, inprocessing, postprocessing, deterministic reranking, optional extras, fit/transform/predict patterns, and fairness-metric validation.
Start here
- Pick the lifecycle stage and class in algorithm-selection.md.
- Follow end-to-end usage patterns in workflows.md.
- Check optional dependency status and install hints in optional-algorithms.md.
- Diagnose install/import, optional-extra, data/API, score, and workflow failures in troubleshooting.md.
- For a base-safe smoke, run reweighing_smoke.py with
--help before --json.
Route neighboring tasks
- Dataset construction,
BinaryLabelDataset fields, group dictionaries, raw dataset wrappers, and metrics: datasets-and-metrics.
aif360.sklearn estimators, pandas protected-attribute indexes, scorers, and sklearn pipelines: sklearn-interface.- MDSS/FACTS subgroup detection or metric text/JSON explainers: detectors-and-explainers.
Operating rules
- Select mitigation by where it acts: preprocessing modifies data before model training; inprocessing trains a mitigated model; postprocessing adjusts already-produced predictions; reranking reorders scored candidates.
- Keep true labels, predicted labels,
scores, protected attributes, favorable/unfavorable labels, and privileged/unprivileged group values aligned across every dataset copy.
- Validate with AIF360 metrics before and after mitigation; a class name is not evidence that fairness improved.
- Optional extras were intentionally not installed in the construction environment. Mark optional workflows as optional/unverified until the user's runtime installs the named extra and passes a small smoke.