ORIGINAL COMPANY-CONTEXT PRACTICE
Validate predictions across geological areas
Original practice scenario inspired by BHP Group Limited's public business context; this is not a reported interview question or a description of its internal systems. A synthetic exploration model predicts a material property from sampled locations. Nearby samples are highly correlated, and some sites are sampled much more densely. How would you evaluate generalization to an unsampled area?
Written for independent practice using company context; this is not a verified question from a company interview.
A REASONED APPROACH
Model answer.
Spatial dependence makes random sample splits optimistic. Evaluate held-out areas that resemble the intended new locations, and keep the prediction claim separate from decisions requiring broader geological evidence.
Work through it
- Define spatial blocks or site groups before splitting, with buffer distances where appropriate. Fit transformations only on the training portion.
- Compare against simple spatial or site-level baselines. Report error by held-out area and sample density, not only a pooled metric dominated by dense sites.
- Assess uncertainty calibration and whether the new area lies outside the training feature range. Describe an abstention or additional-sampling policy for unsupported predictions.
Tradeoffs to discuss
- Larger blocks test stronger extrapolation but leave fewer independent evaluation units. A single spatial split can be unstable, so repeat defensible blocked evaluations.
Common pitfalls
- Nearby train and test points can leak local structure.
- Narrow confidence intervals over correlated rows overstate how much independent evidence the dataset contains.