Shapes
| Shape | Reading | Follow-up |
|---|---|---|
| Monotonic, smooth | steady response | candidate for a monotonic constraint |
| Step / threshold | the model splits near a value (trees) | check the threshold against domain knowledge (policy limits, pricing tiers) |
| U or inverted U | optimum in the middle | confirm with data support at both ends |
| Flat PDP, wide ICE fan | opposite effects cancel | find the separator: an interaction |
| Flat PDP, flat ICE | no effect in this range | consistent with low permutation importance |
| Saw-tooth | overfitting to noise in the feature | regularise; smooth with ALE to compare |
| Jump at the tails | extrapolation or rare outliers | ignore outside the rug; cap the feature |
Heterogeneity
ice_analysis.py distinguishes:
- Opposite effects: some rows go up while others go down. Classic example: a discount raises conversion for price-sensitive customers and has no or a negative effect for loyal ones. The PDP may look flat.
- Subgroups of different strength: everyone moves the same way, by clearly different amounts. Fixture
example: 20 customers lose 0.49 probability over the tenure range, while 10 on monthly contracts lose only
0.16. The separator
contract_monthly(AUC 1.00) names the interaction.
After a separator is found, plot ICE coloured by it, or draw a two-way PDP (--pair). Confirm with SHAP
interaction values for tree models.
Support
PDP values at grid points with no nearby data are model extrapolation. Always draw a rug (the generated
plot does). ice_analysis.py --features flags grid points outside the 5th-95th percentile. Correlated
features shrink the realistic region further: a PDP over tenure at a fixed monthly-charge combination
may never occur, and ALE avoids that.
Misreadings
| Misreading | Correct |
|---|---|
| "Raising X by 10 lowers churn by 5 points" | the model's prediction changes; it says nothing about intervening on X |
| Reading the PDP where the rug is empty | no data there: extrapolation |
| PDP of a feature correlated with others | averages over impossible rows: prefer ALE |
| Uncentred ICE "all parallel" | level differences hide shape; centre the curves |
| Ignoring the response scale | probability effects are not additive; log-odds are |
| Comparing PDP magnitudes across models | depends on each model's calibration; compare on the same scale |
Worked fixture example
$ python ice_analysis.py ice_tenure.csv --features ice_rows.csv
PDP decreases by 0.3805 over the grid (range 0.3805); 100% of steps go the same way
centred ICE: 0 up, 30 down, 0 flat; end-point spread sd 0.1561 (0.41 x the PDP range)
SUBGROUPS: two groups of curves with different strength: 20 rows change by -0.4885, 10 rows by -0.1645
separator: contract_monthly (AUC 1.00: the -0.1645 group has higher contract_monthly)
WARNING: 4 grid value(s) outside the 5th-95th percentile [3, 46.65]: the curve there is extrapolationReading: longer tenure lowers predicted churn for everyone, but much less for monthly-contract customers. Report the two groups separately, and do not read the curve beyond about 47 months.
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.