1. Outcome-side fields
Columns written because the outcome happened.
- Churn:
cancellation_reason,retention_offer_accepted,account_status = 'closed'. - Credit:
collections_flag,charge_off_date,restructured_loan. - Fraud:
chargeback_amount,case_status,analyst_decision. - Healthcare: discharge disposition, billing codes finalised after the stay, "treatment for X" when predicting X. Detection: single-feature AUC, is-missing AUC, outcome-like names (LD001/LD002/LD004).
2. Overwritten current-state fields
A field that is correct today but was different at prediction time: last_login_date, plan, address,
credit_limit, segment recomputed monthly. Joining today's value to a snapshot from six months ago leaks
anything that changed because of the outcome (customers who churned stop logging in; their last_login
is old, but customers who are still active have recent ones even if they were inactive at the snapshot).
Detection: dates later than the snapshot (LD003), signal stronger on old rows (LD008).
3. Aggregates over the future
lifetime_orders, total_tickets, avg_rating computed at extraction time over all history, including
events after the prediction time. Fix: aggregate only events with ts < prediction_time.
4. Proxies of the label
Variables generated by the same process as the label: "number of collection calls" when predicting default; "refund issued" when predicting complaints; a downstream model's score that was trained on the label.
5. Identifiers and ordering
- Autoincrement IDs assigned when a case is opened (cases opened later are more likely to be fraud investigations).
- Files sorted by label; patient IDs encoding the hospital that treats the sickest patients.
- Image/record file names containing the class. Detection: AUC of row index or ID (LD005).
6. Duplicates and near-duplicates across splits
The same customer, device, patient or text appears in train and test (often under multiple snapshots). The model memorises entities. Fix: group-aware splits (GroupKFold), deduplication. See train-test-split-auditor.
7. Target encoding and target statistics in-sample
Category means of the target computed on all rows, including the row being encoded (LK006). With rare levels the encoding is the label.
8. Temporal leakage in validation
Random splits on data where the model predicts the future: training contains later periods than some test rows, so the model learns period effects and uses "future" patterns. Fix: time-ordered splits, gaps equal to the label window when windows overlap (LK007).
9. Preprocessing and selection leakage
Scalers, imputers, PCA, encoders, feature selection or resampling fitted before the split (LK001-LK003). Effect size ranges from negligible (standard scaling, big data) to dramatic (feature selection among thousands of features; SMOTE before split).
10. Leakage through hyperparameter search
Tuning on the test set, or reporting the best of many configurations on the same test data. Fix: nested CV or a final untouched holdout; report the score of the selected configuration on fresh data.
11. Label leakage via joins
Joining a table that is filtered by the label (only churned customers have rows in exit_survey), so the
existence of a joined row encodes the label. Missingness AUC detects it.
12. Competition-style leaks
Leaky metadata (file sizes, timestamps), test-set structure, row order. Useful to recognise; never to exploit in production models.
Real cases worth knowing
| Case | Leak |
|---|---|
| Pneumonia mortality model ranked asthma patients as low risk | asthma patients went straight to ICU (treatment encoded in data) |
Prostate cancer model used PROSSY (surgery) | treatment after diagnosis |
| KDD Cup 2008 breast cancer | patient ID encoded the screening site with different base rates |
| Kaggle competitions where file IDs or timestamps predicted the label | data collection artefacts |
| Credit models using "number of collection calls" | consequence of default |
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.