Structured data

Binary classification

churn, fraud, loan default, lead conversion

A yes/no outcome per customer or transaction, from tabular history.

Probability & Bayes

Odds & log-odds

Maximum likelihood

Information theory

Bias & variance

Hypothesis testing

Foundations

Log-odds and likelihood explain logistic regression; entropy explains tree splits; bias and variance explain why forests and boosting work; hypothesis testing is how you compare two models honestly.

Frame the problem

Frame

Define the unit (customer-month? transaction?) and the horizon. Decide up front what a false positive and a false negative each cost; that sets the metric.

Sourcing & signal

Exploring the data

Hunting for leakage

Labeling & data collection

Cleaning the data

Data & labels

Positives are rare; check they are labelled the same way across time and segments. Churn is usually a derived rule: write it down and test its edge cases. Anything computed after the event leaks: closure dates, chargeback flags, lifetime totals.

Encode & scale

Create features

Reduce & select

Represent

Rolling aggregates over history do most of the work. Every feature must be as-of prediction time. Target-encode high-cardinality ids inside CV folds.

Split the data

Cross-validation

Split

Group by customer. Go temporal if behaviour drifts. Stratify on the rare class.

Dumb baseline

Logistic regression

Decision trees

Bagging & random forests

Gradient boosting

Variants worth knowing

Credit risk is this playbook plus regulation: reason codes, monotonic constraints, reject inference. Fraud rings are a graph problem: transaction-graph features, or a GNN. Networks on the table itself only with text or image columns, or very large data.

Interpretability

Graph neural networks

Networks on tabular data

Model

Majority class, then logistic regression on a handful of features, then gradient boosting. Stop there unless you have a reason.

Loss functions

Hyperparameter tuning

Handle imbalance

Over- & underfitting

Ensembling

Calibration

Train

Class weights before SMOTE. Early stopping on a validation fold. Average a few seeds before reaching for stacking. Calibrate if the score feeds a threshold.

Classification metrics

Interpretability

Error analysis & significance

Evaluate offline

PR-AUC, not accuracy. Slice by tenure and segment. Pick the threshold from the cost of errors in step 1, not 0.5.

A/B testing & interleaving

Evaluate online

If the score drives an intervention such as a retention offer, test the intervention, not the score. See causal inference & uplift.

Features at serving time

Serving & release

Monitor & retrain

Ship & monitor

As-of features must be recomputed identically online. Acting on flagged users changes future labels; plan for the feedback loop.

React Flow