AIX360

Use this repository skill for AI Explainability 360 (aix360) workflows over tabular, text, image, and time-series data. It covers package version 0.3.0, including modern CPU-usable routes and explicit boundaries around historical TensorFlow/Keras and other optional dependency stacks.

AIX360 is a toolkit, not one universal explainer. First identify:

1. whether the artifact to explain is data, an already-trained black-box model, or a directly interpretable model to fit;
2. whether the requested result is local, global, counterfactual, certified, or a quality metric;
3. the input domain and model-callable output shape;
4. constraints on actions, downloads, hardware, latency, and optional packages.

Installation and inspection

Use an isolated environment because AIX360 extras pin mutually incompatible versions for some algorithm families.

python -m pip install aix360
python -c "import aix360; from importlib.metadata import version; print(version('aix360'))"

Install only the extra for the chosen workflow, for example:

python -m pip install 'aix360[lime]'
python -m pip install 'aix360[tslime]'
python -m pip install 'aix360[rbm]'

Do not install every extra into one environment. In particular, CEM/ProfWeight and historical SHAP paths pin TensorFlow 1.14 with Keras 2.3.1, whereas nearest-neighbor contrastive pins TensorFlow 2.9.3; keep incompatible families in separate environments. Read installation and runtime troubleshooting before changing versions.

Run the bundled environment diagnostic to inspect the base package and selected optional modules without downloading data or models.

Route by task

Local black-box attribution and examples

Read local-black-box for LIME, SHAP, Grouped Conditional Expectation, nearest-neighbor contrastive examples, faithfulness, monotonicity, prediction-callable contracts, feature names, and local explanation output validation.

Typical signals: explain_instance, predict_proba, local feature weights, SHAP values, tabular/text/image attribution, exemplar/nearest-neighbor explanation, or local metric debugging.

Counterfactuals, recourse, certification, and matching

Read counterfactual-and-certification for CEM/CEM-MAF pertinent positives and negatives, Ecertify trust regions, GLANCE recourse/action costs, and order-constrained optimal-transport matching.

Typical signals: target class, actionable or immutable features, feature bounds, recourse, robustness certificate, perturbation budget, OTMatchingExplainer, or legacy CEM model setup.

Directly interpretable models, rules, and prototypes

Read interpretable-models for ProtoDash, Boolean/linear rule models, RIPPER/TRXF, interpretable model differencing, teaching explanations, and optional CoFrNet, DIPVAE, and ProfWeight workflows.

Typical signals: prototype selection, FeatureBinarizer, BRCG/GLRM, rule induction, model comparison, explanation labels, directly interpretable training, solver errors, graph export, or rule serialization.

Time-series explanations

Read time-series for TSICE, TSLime, and TSSaliency over univariate or multivariate histories, including forecast lookahead, relevant history, exogenous variables, perturbation windows, data shapes, and numeric-versus-plot output.

Typical signals: temporal attribution, integrated gradients, local surrogate, forecast window, time axis, feature axis, perturbation count, or exogenous series alignment.

Datasets, preprocessing, and explanation metrics

Read datasets-and-metrics for AIX360 dataset constructors, local data layout, offline checks, preprocessing, and Faithfulness/Monotonicity metrics.

Typical signals: HELOC, COMPAS, CDC, MEPS, Ford, Sunspots, CIFAR, MNIST, CelebA, e-SNLI, missing dataset paths, downloads, coefficient alignment, or explanation-quality evaluation.

Cross-route decisions

• Use dataset and metric guidance as support for any algorithm route, but keep explainer construction with the algorithm-owning sub-skill.
• Prefer a callable whose batch input and output shape are explicit; many local explainers fail because a classifier returns labels instead of probabilities or a time-series forecaster drops its batch axis.
• Treat notebook-scale image training, remote datasets, pretrained weights, and graph rendering as opt-in operations. The bundled skill defaults to tiny, local, deterministic checks.
• Verify whether an explanation is local/global and post-hoc/direct before comparing methods. Their outputs are not interchangeable.
• A successful import aix360 proves only the base package. Import the selected algorithm module and run a tiny fixture before trusting an optional route.

Read API and method overview when the user names an algorithm but not its route. Read repository provenance before deciding whether this skill is stale for another checkout.