Protocol
- Fix the evaluation first: split strategy (time / group / stratified), primary metric at the operating point, secondary metrics, and the final untouched holdout.
- Shortlist 3-5 models: logistic baseline, 1-2 GBMs, one model per hard constraint (EBM if explanations are mandatory), and the incumbent production model.
- Round 1: defaults on identical folds. Purpose: find the family that fits the data, not the winner.
- Round 2: equal tuning budget for the top two (same number of trials, same search method, same inner CV), nested inside the outer folds or followed by a fresh holdout.
- Compare with paired statistics on per-fold scores (corrected resampled t-test, Wilcoxon) and bootstrap CIs on the holdout.
- Decide with the decision rule (simplest within noise that meets constraints).
- Confirm on the untouched holdout once.
Folds
| Data | Splitter | Notes |
|---|---|---|
| i.i.d. rows | StratifiedKFold(5, shuffle=True, random_state=0), repeated 3x for small data | report mean ± std |
| Repeated entities | StratifiedGroupKFold | the same entity never in train and test |
| Forward prediction | TimeSeriesSplit or rolling origin with a gap | the most recent fold matters most |
Precompute fold indices once and pass the same list to every model (the generated script does this).
Metrics and costs
Record per fold: primary metric, secondary metrics, fit seconds, prediction latency per 1k rows, model size (pickle bytes). Latency and size often decide between two models within noise.
Budgets
| Model | Round-2 search |
|---|---|
| Logistic | C on a log grid (10-20 values), penalty type |
| GBMs | 50-100 random/Bayesian trials over learning rate, leaves/depth, min child size, subsampling, regularisation; early stopping inside each fold |
| RF | max_features, min_samples_leaf (10-20 trials) |
| EBM | interactions, max_bins (5-10 trials) |
Give every finalist the same number of trials; report the budget.
Time-based bake-off
For forecasting-like classification (churn, fraud), add a "by period" view: train on months 1..k, test on month k+1, for several k. A model that wins on average but loses on the latest periods is a worse choice.
Common pitfalls
| Pitfall | Effect |
|---|---|
| Different folds per model | differences reflect folds, not models |
| Tuning only the favourite | biased comparison |
| Selecting the winner on the test set and reporting its test score | optimistic |
| Comparing means without fold-level pairing | over- or under-states significance |
| Ignoring cost columns | a 0.002 AUC gain that doubles latency |
| Using impurity importances to compare models | not comparable across model types |
Report template
Bake-off: churn_30d, 400 customers, StratifiedKFold(5, shuffle, seed 0), primary PR-AUC
| model | PR-AUC mean ± sd | ROC-AUC | log loss | fit s | ms/1k rows |
| logreg | 0.621 ± 0.095 | 0.799 | 0.400 | 0.01 | 33 |
| ebm | 0.605 ± 0.114 | 0.787 | 0.413 | 2.7 | 8 |
| catboost | 0.557 ± 0.070 | 0.763 | 0.515 | 0.56 | 7 |
Paired test logreg vs ebm: difference +0.016, corrected t-test p = 0.6 (no evidence of a difference).
Decision: logistic regression (simplest, fastest, best calibrated); revisit GBMs when data > 5k rows.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).
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.
counterfactual-explanation-generator
Generate counterfactual explanations and algorithmic recourse: the smallest realistic change to a row's features that changes the model's decision. Solves exact minimal-cost recourse for linear and logistic models (MAD-scaled L1 cost, immutable features, increase/decrease directions, bounds, integer and binary features, diverse alternatives, features every counterfactual must change, and a no-recourse finding when immutable features decide), generates DiCE counterfactuals for any scikit-learn model with constraints turned into permitted ranges and every result re-predicted, and validates counterfactuals from any method for constraint violations, implausible values, off-manifold rows, broken derived-feature relations and failure to flip the prediction.