Why Powerful ML Is Deceptively Easy — Part 2

Powerful machine learning models for spatial prediction can be deceptively convincing when the evaluation setup is flawed by issues like spatial dependence and temporal leakage. Several common pitfalls, such as the "Proximity and Persistence Trap," cause models to appear more generalizable than they are by exploiting similarities in nearby or repeated data points. Other issues include the "Coverage Illusion," where performance is skewed by dense areas, and "Geographical bias," where models may inadvertently reinforce historical segregation. Designing a robust evaluation framework that accounts for these spatial structures, using methods like temporal-spatial holdouts, is more critical than simply fitting a complex model.