3.1.1. Mapping Business Problems to ML Algorithms

💡 First Principle: ML problems fall into a small number of categories—classification, regression, clustering, anomaly detection, recommendation, and forecasting—and each category has algorithms that are well-suited to it. The first step in model selection is always identifying the problem category, not the algorithm.

Interpretability considerations: Some business contexts require model interpretability—you need to explain why the model made a prediction. Linear models and tree-based models are inherently interpretable. Neural networks are not. If a question mentions "regulatory requirements," "explain decisions to customers," or "audit trail for predictions," interpretability should drive your model choice toward simpler, explainable models—or toward explaining a complex model with SHAP values (the open-source SHAP library, which is also the engine inside SageMaker Clarify).

⚠️ Exam Trap: When a question describes a "prediction problem," determine whether it's classification or regression before selecting an algorithm. "Predict whether a customer will churn" is classification (binary output). "Predict how much a customer will spend" is regression (continuous output). The same algorithm family (XGBoost, Linear Learner) can do both, but the configuration differs—and the evaluation metrics differ entirely.

Reflection Question: A retail company wants to predict which products a customer will buy next, estimate their total quarterly spending, and identify fraudulent return patterns. Classify each as a problem type and recommend an algorithm.