A steadier shortlist for nickel

Seongin Ahn’s team compared XGBoost with a positive-unlabelled ensemble for nickel prospectivity in Western Australia’s Yilgarn Craton. The 2024 dataset covered 70,200 five-kilometre grid cells, 818 containing known deposits.[1]

The underlying records describe an explored landscape. Western Australia’s MINEDEX service includes mines, deposits and prospects, with geology, operating status, project history, resource estimates and production data. These are different kinds of evidence about a location. The database can be searched spatially through GeoVIEW.WA and downloaded for analysis. A useful follow-up to any regional shortlist is therefore to inspect what the records actually contain: a mineralised site, an estimate of its resource, and a producing mine establish different things.[2]

The ensemble method draws on work by Fantine Mordelet and Jean-Philippe Vert. Their bagging approach keeps the known positive examples while repeatedly changing the sample drawn from the unlabelled pool, then averages the resulting classifiers. Hidden positives appear in different proportions across those samples. A single unlucky selection can distort a classifier; aggregation reduces dependence on that selection. The mechanism addresses variability created by the available training examples. The ensemble still learns from the same available evidence.[3]

Overall precision–recall performance was virtually unchanged. Among each model’s highest-scoring 1% of cells, however, the median relative interquartile spread fell from 48% with standard XGBoost to 18% with the ensemble. This compares the middle half of resampled predictions with their mean.[1]

Both models concentrated about 55% of known positive cells within their three highest-priority zones, covering roughly 5% of the area. Stability changed the priority tiers. New discoveries and drilling-cost savings were not tested.[1]

Report an error or suggest a correction