Pillar 1 · Model Selection — family deep dive

Types of Classic ML

No neural networks here — and on structured business tables these methods still win on accuracy, cost and explainability. The types split along one question first: do you have labeled outcomes to learn from? Everything else follows from that.

Picking the type

The classic-ML decision tree

Do you have labeled outcomes in the data? yes — supervised Predict a number? price · demand · risk score Regression linear · XGBoost regressor Predict a category? churn yes/no · fraud · segment Classification logistic · LightGBM · forest no — unsupervised Group similar records? customer segments Clustering k-means · DBSCAN Spot unusual records? fraud · faults · outliers Anomaly detection Isolation Forest · LOF Data is a time-ordered series? demand, traffic, sensor readings → forecasting models Forecasting Prophet · ARIMA · boosting
Supervised = learn from labeled history · unsupervised = find structure without labels · forecasting = the time axis is the structure.

The types, with named models

TypeNamed models / librariesChoose it whenTrade-off
Linear / logistic regressionscikit-learn, statsmodels; credit scorecardsBaselines, pricing, regulated risk scoring where every coefficient must be defensibleUltimate explainability; misses non-linear patterns
Decision treeCART (scikit-learn)You need human-readable if/then rules for the businessReadable but weak alone — usually a building block
Random forestscikit-learn RandomForestRobust "just works" model with little tuningSolid, rarely best; slower to score than boosting
Gradient boostingXGBoost, LightGBM, CatBoostThe tabular default — churn, fraud, pricing, ranking; wins most tabular benchmarksNeeds feature engineering + retraining cadence for drift
SVMLIBSVM / scikit-learn SVCSmall, high-dimensional datasets (mostly legacy today)Doesn't scale to big data; boosting usually beats it
Clusteringk-means, DBSCAN, hierarchical (scikit-learn)Customer segmentation, store grouping — no labels needed"Right" cluster count is a judgment call; validate with the business
Anomaly detectionIsolation Forest, One-Class SVM, LOFFraud/fault detection when labeled fraud examples are scarceFlags unusual, not necessarily bad — needs human triage
Time-series forecastingProphet, ARIMA/SARIMA, LightGBM with lag features; deep options: N-BEATS, TFTDemand, capacity, cash-flow forecastingRegime changes break history-based models — monitor and refit
Default play for tabular data: start with gradient boosting (XGBoost/LightGBM) plus a linear baseline. If boosting doesn't clearly beat the baseline, your features — not your model — are the problem. Deep learning enters tabular problems only with very large data or mixed unstructured inputs.
Why this family is an architect's friend: CPU-only inference (no GPU line item), millisecond scoring, mature explainability tooling (SHAP feature importance) for auditors — and your existing data warehouse is the feature store.
← GAN typesStyleGAN, CycleGAN, CTGAN