Wording
"Based on the information we used, any one of the following would likely have led to a different outcome:
- reducing your number of open support tickets from 10 to 7 and logging in within the next two weeks, or
- reducing your number of open support tickets to 5. These are not guarantees. Our assessment can also change as other information changes."
- Give options, ordered by effort, in real units and time frames.
- Avoid technical terms (MAD, log-odds) and model internals.
- Do not present a single counterfactual as "the reason". Reasons come from attribution (SHAP reason codes).
- Do not suggest changing immutable or protected characteristics, or their proxies.
- Re-check before sending: the counterfactual must still flip the current model version.
Regulation (orientation, not legal advice)
| Context | Relevance |
|---|---|
| US credit (ECOA / Regulation B adverse-action notices, FCRA) | principal reasons for denial are required. Counterfactuals can complement them, but do not replace them |
| EU GDPR Art. 22 and recital 71 | safeguards for solely automated decisions, including the right to contest. Counterfactuals are widely discussed as a way to provide meaningful information (Wachter, Mittelstadt and Russell 2017) |
| EU AI Act (high-risk systems) | transparency and human oversight obligations; recourse supports both |
Have legal and compliance review the method and the wording before customer-facing use.
Fairness of recourse
Equal accuracy does not imply equal recourse. Audit across groups:
- the share of negatively decided people with no recourse (immutable features decide)
- the cost of the cheapest recourse (median MAD units) per group
- essential features per group (e.g. income is essential for one group and not another)
# loop linear_counterfactual.py over rejected rows and aggregate by group
for row in rejected/*.json; do
python linear_counterfactual.py model.json "$row" --threshold 0.5 --json
done > recourse.jsonlLarge gaps in recourse cost (Ustun et al. 2019; Gupta et al. 2019 "equalizing recourse") belong in the model card (model-card-writer) and may need model changes (fewer immutable drivers, monotonic constraints).
Gaming and robustness
- Counterfactuals reveal the model's decision boundary. Customers may change proxies without changing the underlying risk ("game" the model). Prefer recourse through features that are causally related to the outcome.
- A counterfactual valid for today's model can be invalid after a retrain. Recompute it, and avoid making commitments ("if you do X, you will be approved") unless you will honour them.
- Prefer robust counterfactuals: a small margin beyond the threshold (
--margin, or a target probability below the threshold) survives small model changes.
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.