Learning from Tabular Data: The Tree Era

Distribution shift: the world changes between training and deployment (as in the fraud example above).
Hidden dependencies: two rows from the same customer are not independent.
Sample selection bias: the data you collect isn't drawn from the population you serve.

A deep dive into why tree-based models dominate tabular machine learning, from the inductive bias war between trees and neural networks through decision trees, random forests, gradient boosting, and XGBoost.