The challenge begins with how exploration data are collected. Traditional machine learning approaches to mineral prospectivity mapping treat every grid cell containing a known deposit as a positive example and every other cell as a negative one. But that assumption is geologically naive. Exploration has historically concentrated on accessible, well-mapped, and heavily prospected terrain, meaning that remote or poorly exposed areas may host substantial mineralization that simply has never been found. Treating those unexplored cells as confirmed barren ground biases the model against geological signatures resembling undiscovered deposits. The research team, led by Seongin Ahn of the Korea Institute of Geoscience and Mineral Resources, adopted positive-unlabeled learning, a statistical framework that treats the unlabeled background as a mixture of true negatives and unobserved positives rather than a homogeneous barren class.

Their implementation wraps the popular gradient-boosting algorithm XGBoost in a bagging procedure designed for positive-unlabeled data. Instead of training one model on the full dataset, the BaggingPU-XGBoost variant trains many base learners, each on a bootstrap subsample of the unlabeled cells combined with all known deposit cells, and averages their predictions. Because each subsample contains a different provisional negative set, the influence of any hidden positives accidentally labeled as negatives is diluted across the ensemble. The approach follows methods developed in computational biology and has been applied to prospectivity mapping before, but the Yilgarn study couples it with several other techniques rarely used together in exploration geoscience.

One of those techniques addresses a second, subtler source of bias: spatial autocorrelation. Geological units, geophysical responses, and mineral occurrences are strongly correlated over distance, so randomly splitting data into training and test sets places near-identical neighboring cells on both sides, inflating apparent model skill. The researchers instead used nested spatial block cross-validation. An empirical variogram of the deposit labels showed spatial autocorrelation dying out at roughly 31.5 kilometers, so they partitioned the study area into 702 blocks of 50 kilometers per side, well beyond the autocorrelation range, and assigned blocks to five folds. Within each outer training set, an inner fivefold spatial loop handled hyperparameter tuning, keeping the outer test folds completely untouched until final evaluation. The result is an honest estimate of how well the model transfers to genuinely new ground rather than to the neighborhood of deposits it has already seen.

The study area itself is demanding. The Yilgarn Craton is an Archean granite-greenstone terrane hosting world-class komatiite-hosted nickel sulfide deposits such as Kambalda, Perseverance, and Mount Keith, formed during peak ultramafic volcanism around 2.7 billion years ago. The researchers assembled 21 predictors from public datasets of the Geological Survey of Western Australia and Geoscience Australia, including distances to ultramafic units and major crustal structures, high-resolution magnetic and radiometric grids, gravity data, lithosphere thickness estimates, and multi-scale edge features derived from potential field data. Everything was harmonized onto a grid of 70,200 cells at 5-kilometer resolution, of which just 818 cells, or 1.17 percent, contain at least one known nickel deposit. That severe class imbalance shaped the entire analysis, from the choice of precision-recall area as the primary metric to the interpretation of predicted probabilities as relative favorability scores rather than literal odds of striking ore.

On the spatially independent outer test folds, the two models performed almost identically in aggregate: PR-AUC of 0.195 for XGBoost and 0.194 for BaggingPU-XGBoost, with both achieving ROC-AUC of 0.911. The positive-unlabeled approach showed an advantage during inner model selection, but that edge vanished on held-out ground, possibly because the mature exploration coverage of the Yilgarn has already located most large occurrences, leaving fewer hidden positives in the unlabeled pool than a frontier province would contain. The fold-level variation was striking, with outer PR-AUC ranging from about 0.11 to 0.32, a reminder that transferability across spatially distinct blocks varies enormously and that a single aggregate number can mask real geographic variability in model reliability.

The real innovation lies in how the framework converts raw probability into decision-ready guidance. Fifty bootstrap models were trained per fold, and the spread of their predictions was summarized by a relative interquartile range, a scale-independent measure of how much the ensemble disagrees about each cell. Cross-conformal calibration then translated the bootstrap mean probability into a coverage-controlled prediction set: using out-of-fold predictions and Platt scaling, the researchers set a threshold guaranteeing at least 85 percent of known deposits would fall inside the conformal set. In practice, empirical coverage exceeded the target, at roughly 91 percent for both models, while the conformal filter excluded more than half of the study area from immediate consideration. In exploration terms, the exclusion removed about one in ten known deposits while screening out a vast swath of low-support ground.

Combining the bootstrap mean probability, the uncertainty measure, and conformal membership, the framework assigns every cell to one of five zones, from immediate targets down to conformally excluded areas. Here the two models diverged dramatically despite their identical aggregate scores. In the top 1 percent of predicted probabilities, the median relative interquartile range was 18 percent for BaggingPU-XGBoost versus 48 percent for plain XGBoost, meaning the positive-unlabeled ensemble produced far more reproducible high-probability predictions. Under the 30 percent uncertainty cut-off, BaggingPU-XGBoost assigned 0.68 percent of the study area to the highest-priority Immediate zone, capturing 17.25 percent of known deposits there, while XGBoost managed only 0.73 percent of deposits in its tiny Immediate zone. The combined Immediate and Priority zones captured 33 percent of positive cells for BaggingPU-XGBoost against 23 percent for XGBoost, a redistribution driven by prediction stability rather than any difference in the underlying geological signal.

Interpretability analysis confirmed that both models learned similar regional controls, with proximity to ultramafic sources emerging as the dominant predictor, consistent with the established understanding that komatiite-hosted nickel sulfide mineralization is governed by deep ultramafic plumbing and craton-margin structures. The authors are careful about what their uncertainty measure does and does not capture: the bootstrap dispersion quantifies only model instability under resampling, not uncertainty from incomplete data, missing predictors, or unrepresented geological processes. Within the high-probability decile, low dispersion largely tracks high probability itself, so the uncertainty signal functions as a reproducibility filter that triages already-favored cells into immediate follow-up versus additional-evidence categories, rather than as an independent discovery tool. Sensitivity tests showed the zone output is robust to moderate changes in the tier boundaries and the uncertainty cut-off.

The framework’s limitations are acknowledged plainly. Validation covered a single craton, and the conformal guarantee remains approximate because spatial autocorrelation violates the strict exchangeability that conformal prediction assumes, even with 50-kilometer blocks. The thresholds are heuristic operating points, not universal rules, and would need recalibration for greenfield provinces where the positive-unlabeled formulation might deliver larger benefits. Whether the prioritization actually saves money requires prospective drilling with explicit cost accounting, which no retrospective study can provide. Still, the core message resonates across a discipline racing to explore under deep cover as near-surface deposits are exhausted: a probability map that cannot say how confident it is, or where its own support ends, is only half a map. By fusing positive-unlabeled learning, spatially honest validation, bootstrap stability, and conformal screening into a single decision layer, this work offers exploration geologists something genuinely new, a targeting product that knows the limits of its own knowledge and says so, cell by cell, before anyone commits a drill rig.

Subject of Research: Uncertainty-aware machine learning for nickel mineral prospectivity mapping in the Yilgarn Craton

Article Title: Uncertainty-Aware Mineral Prospectivity Mapping with Positive-Unlabeled Learning and Conformal Calibration: Ni Mineralization in the Yilgarn Craton

Article References: Ahn, S., Jo, H., Kwon, J., Park, G., & Lee, S.-H. (2026). Uncertainty-Aware Mineral Prospectivity Mapping with Positive-Unlabeled Learning and Conformal Calibration: Ni Mineralization in the Yilgarn Craton. Natural Resources Research. https://doi.org/10.1007/s11053-026-10778-w

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10778-w

Keywords: mineral prospectivity mapping, positive-unlabeled learning, conformal prediction, uncertainty quantification, spatial cross-validation, XGBoost, nickel deposits, Yilgarn Craton, komatiite-hosted mineralization, exploration targeting, bootstrap ensembles, machine learning

The core difficulty is easy to state but hard to solve. In classical binary classification, an algorithm sees both labeled positives and labeled negatives and learns a decision boundary between them. In PU learning, negative labels simply do not exist. Instead, the algorithm receives a small set of confirmed positives and a large unlabeled pool. Because the unlabeled pool is dominated by negatives but contaminated with positives, treating every unlabeled example as negative injects label noise into training. Traditional approaches either try to identify reliable negatives from the unlabeled set before training, or they incorporate assumptions about the class prior, the underlying proportion of positive examples in the data. Both strategies become brittle when the dataset is severely imbalanced, meaning positives may account for only one to a few percent of all examples.

Zavitsanos and Paliouras observed that in imbalanced PU data, the mathematics of existing risk estimators works against the practitioner. The estimators used by state-of-the-art methods, such as the unbiased PU estimator (uPU) and its non-negative successor (nnPU), contain a term weighted by the positive class prior. When that prior is tiny, the contribution of the few labeled positives nearly vanishes, and the training objective is dominated by the unlabeled data. The model therefore learns to be excellent at recognizing negatives, and mediocre at catching the rare positives that matter most. Matters worsen under the Probabilistic Gap assumption, a realistic refinement of labeling theory in which positive examples that resemble negatives are the least likely to have been labeled. The very positives a model most needs to learn from are the ones most likely to be hidden in the unlabeled pool.

The researchers’ answer is a new empirical risk estimator they call iFPU, for imbalanced focused PU learning. The key ingredient is focal loss, a function originally developed in computer vision for dense object detection, where positive targets are similarly swamped by an enormous number of easy negatives. Focal loss reshapes the standard cross-entropy objective by multiplying each example’s loss by a factor that shrinks as the model’s confidence in the correct answer grows. Easy examples, which the model already classifies correctly with high probability, contribute almost nothing to the gradient. Hard examples, sitting near the decision boundary, receive exponentially amplified weight, controlled by a focusing parameter gamma. The result is that training concentrates its capacity on exactly the boundary cases that define the Probabilistic Gap.

Incorporating focal loss into a PU risk estimator is technically delicate. The authors derive a non-negative risk estimator that combines three terms: the focal loss of the labeled positives treated as positive, an unbiased correction subtracting the focal loss the positives would incur if treated as negative, and the focal loss of the unlabeled pool treated as negative. A max operator clamps the correction at zero to prevent the notorious problem of negative empirical risk, which causes overfitting in neural networks trained with earlier unbiased estimators. When the correction term goes negative within a training mini-batch, the method performs a step of gradient ascent instead, deliberately nudging the model away from overfitting that batch. The whole procedure slots into standard stochastic optimization, requires no preprocessing, resampling, or manipulation of the data, and can be attached to essentially any classifier trained by cost minimization, from multilayer perceptrons to gradient-boosted trees.

The authors also supply a rigorous theoretical analysis. They prove that the population-level iFPU risk is identical to the focal classification risk under full supervision, for any labeling mechanism, by virtue of a mixture identity between the unlabeled distribution and the class-conditional densities. They further characterize the bias that appears when labeled positives are not selected completely at random, showing that this bias depends only on the labeling propensity and is orthogonal to the choice of loss function. Under the SCAR assumption, they establish an estimation error bound that vanishes at the standard statistical rate, proportional to the inverse square root of the number of labeled positives and unlabeled examples combined, using Rademacher complexity tools. Crucially, they show focal loss is Lipschitz continuous and bounded on a restricted score range, which makes these guarantees possible. Although focal loss itself introduces a deliberate bias by reweighting errors, it remains classification-calibrated, meaning the Bayes-optimal classifier under focal loss coincides with the Bayes-optimal classifier under zero-one loss.

Empirically, the method was tested on 14 publicly available benchmark datasets for imbalanced binary classification, spanning positive rates from roughly 1 to 14 percent. The experimental design was deliberately demanding: positive examples were progressively hidden in the unlabeled pool at rates of 25, 50, and 75 percent, under both the SCAR and the more realistic SAR labeling assumptions, generating 840 experimental runs in total. Notably, the authors avoided hyperparameter tuning altogether, using the recommended default gamma of 3, because tuning in PU settings is itself fraught with assumptions due to the absence of negatively labeled validation data. Under SCAR, iFPU outperformed the neural risk estimators uPU, nnPU, and i-NNPU, and when paired with an XGBoost classifier it matched or exceeded strong competitors including the two-step NNIF anomaly-detection method, PU Hellinger Decision Trees, the label-bias estimation method LBE, and the SAREM expectation-maximization framework, trailing only slightly behind the PU Hellinger Random Forest ensemble.

The picture shifts decisively in favor of iFPU under the more realistic SAR assumption, where positives resembling negatives are less likely to be labeled. Here the competing methods degrade noticeably, while iFPU maintains its performance, and in the hardest scenario, with only 25 percent of positives labeled, it ranks first overall, surpassing the PU Hellinger Random Forest by five percentage points in PR-AUC. Statistical tests confirmed significant differences among methods, and paired comparisons showed iFPU significantly outperforming all alternatives in the most challenging configuration. A sensitivity analysis demonstrated that the method remains robust even when the class prior is misspecified by factors of two or four in either direction, degrading meaningfully only when the prior is severely underestimated. On the three most imbalanced datasets, Cover, Poker, and Satellite, iFPU showed clearly higher mean and median PR-AUC than its strongest rival.

To demonstrate real-world value, the researchers applied their method to financial misstatement detection, a problem where PU data arise naturally. Auditors and regulators typically discover accounting misstatements years after reports are filed, and often only a fraction of misstatements have been identified when a model is trained. Using data on publicly traded US companies spanning 2000 to 2014, with 47,086 firm-year records described by 28 financial indices and derived accounting features, the team simulated realistic detection delays in which approximately 40 percent of positive training labels were missing at training time. Misstatements, whether deliberately concealed frauds or subtle errors resembling normal accounts, fit the Probabilistic Gap assumption perfectly. Models built on a TabTransformer architecture with gated MLP modules and equipped with the calibrated iFPU risk achieved R-precision scores roughly three times higher than prior baselines such as RUSBoost, and outperformed both earlier specialized models and the PU Hellinger Random Forest, setting a new state of the art in this application.

The significance of this work extends beyond any single benchmark. By combining a principled risk-estimation framework with a loss function engineered for imbalance, the authors show that PU learning can be made practical in exactly the conditions that dominate high-stakes applications, from disease gene identification to fraud detection, where positives are rare, partially labeled, and deceptively similar to negatives. The method’s plug-and-play compatibility with modern classifiers, its robustness to prior misspecification, and its theoretical grounding distinguish it from heuristic preprocessing pipelines. The authors point to extensions toward semi-supervised and multi-class settings, integration with pre-trained tabular foundation models such as TabPFN, and output calibration via temperature scaling as promising future directions. For now, iFPU offers practitioners a rare commodity in weakly supervised machine learning: a method whose assumptions match reality rather than convenience.

Subject of Research: Positive-unlabeled machine learning from highly imbalanced datasets

Article Title: Focused PU learning from imbalanced data

Article References: Zavitsanos, E., & Paliouras, G. (2026). Focused PU learning from imbalanced data. Data Mining and Knowledge Discovery, 40(6), Article 101. https://doi.org/10.1007/s10618-026-01264-1

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01264-1

Keywords: PU learning, imbalanced classification, weakly supervised learning, focal loss, risk estimation, machine learning, fraud detection, financial misstatement detection, XGBoost, SCAR assumption, SAR assumption, data mining

Mineral prospectivity mapping, or MPM, is the science of turning scattered geological evidence into predictive maps that show where undiscovered ore bodies are most likely to lie. Over the past decade, machine learning has transformed this discipline: algorithms such as random forests, support vector machines, and deep neural networks can digest vast layers of geological, geophysical, and geochemical data and highlight subtle patterns invisible to human analysts. Yet the field has a persistent weakness. Purely data-driven models treat the Earth as a statistical puzzle, ignoring the physical processes that actually create ore deposits. The result can be models that score well on paper but contradict what geologists know about how mineralizing fluids move, react, and precipitate metals. The new study was designed specifically to close that gap between statistical power and geological realism.

The researchers’ solution is a knowledge–data dual-driven framework built around the mineral system concept, which describes ore formation as a chain of critical ingredients: a metal and fluid source, pathways that focus fluid flow, and traps where gold precipitates. For orogenic gold, structural architecture is paramount. Gold-bearing fluids rise along faults, shear zones, and fracture networks, so the density and geometry of geological lineaments—linear features visible in remote sensing and structural data—serve as a first-order proxy for fluid focusing. The team embedded this knowledge at two distinct levels of the modeling workflow. First, during feature engineering, they prioritized structural characteristics, especially lineament density, as the dominant controlling evidence layers alongside stratigraphic and geochemical information. Second, and perhaps more innovatively, they used geological criteria to shape the training data itself through a positive-unlabeled learning strategy.

The positive-unlabeled, or PU, approach addresses one of the most stubborn problems in exploration machine learning. Models need both positive examples, meaning known deposits, and negative examples, meaning genuinely barren ground. But in reality, geologists rarely know that an area is barren; they simply have not looked there yet. Randomly sampling negatives therefore injects systematic bias, because many supposedly negative cells may actually host undiscovered gold. The Guangxi team instead restricted the selection of negative samples using structural, stratigraphic, and geochemical criteria, excluding only locations that confidently fail multiple ore-control tests. This knowledge-guided sampling reduces label noise, enhances geological representativeness, and produces training sets that better reflect the true mineral system rather than the history of past exploration campaigns.

With the framework in place, the researchers systematically evaluated four representative machine learning algorithms: random forest, support vector machine, deep neural network, and deep forest. Deep forest, an ensemble method that stacks cascades of decision-tree forests rather than relying on differentiable neural layers, emerged as the clear winner, delivering the best prediction accuracy, robustness, and generalization. The comparison matters because it demonstrates that architectural sophistication alone does not guarantee exploration success; the interplay between algorithm design and the structure of geologically informed input data determines performance. Deep forest’s hierarchical, non-parametric character appears particularly well suited to the nonlinear, threshold-like relationships that govern whether a fault intersection becomes a world-class gold deposit or an empty structure.

Interpretability was a central design goal rather than an afterthought. Using feature importance analysis and SHapley additive explanations, a game-theoretic technique that quantifies each variable’s contribution to individual predictions, the team showed that the control sequence is consistent across models: structural parameters dominate, followed by geochemical anomalies and then stratigraphic units. This hierarchy mirrors what field geologists have long argued about orogenic gold systems, in which focused fluid flux along structures is the decisive ingredient, with favorable host rocks and chemical traps playing supporting roles. When a black-box model’s internal logic aligns with independent geological understanding, confidence in its predictions rises sharply, and the model becomes a tool for testing scientific hypotheses rather than merely generating colored maps.

To test whether the framework generalizes beyond its training region, the researchers applied it independently to the Youjiang Basin and the Qin–Hang Belt, two domains in southern China that both host orogenic gold deposits but differ markedly in geological character. The verdict was nuanced and scientifically valuable. The structural component of the model proved broadly transferable, confirming that fault-focused fluid flow is a near-universal control on orogenic gold. However, lithological traps and geochemical signatures varied between the two belts and limited full model portability. In other words, the skeleton of the mineral system travels well, but its flesh is regionally specific. This finding offers practical guidance for exploration companies: structural prospectivity layers can be exported across terranes with reasonable confidence, while geochemical and lithological evidence layers should be rebuilt locally before being trusted.

The practical payoff of the study is a set of newly identified prospective targets in both the Youjiang Basin and the Qin–Hang Belt, providing clear direction for the next stage of gold exploration in Guangxi and its neighboring provinces. The Youjiang Basin, straddling the China–Vietnam border, is famous for Carlin-style and orogenic gold systems hosted in folded sedimentary rocks, while the Qin–Hang Belt records a long history of tectonic activity along the southern margin of the South China block. New targets in such mature provinces are increasingly hard to find near the surface, and the dual-driven model’s ability to integrate deep structural architecture with surface geochemistry offers a way to see through cover and complexity. Each mapped high-probability cell represents a hypothesis that can now be tested with targeted geophysics and drilling, converting statistical prediction into economic action.

Beyond Guangxi, the study carries a broader message for the future of exploration science. As easily discovered deposits become scarce and the global demand for gold and critical minerals intensifies, the mining industry is turning to artificial intelligence to squeeze more value from existing data. But the Guangxi work argues that the winning formula is not raw algorithmic horsepower. It is the disciplined marriage of machine learning with mineral systems knowledge—embedding structural controls into feature design, constraining sampling with geological criteria, and validating interpretability against process understanding. The authors describe their workflow as a practical template for prospectivity mapping in structurally complex regions worldwide, and the transferability experiments suggest it can be adapted wherever orogenic gold, or indeed other structurally controlled deposit types, are sought. If the newly flagged targets deliver, the study may mark a turning point in how geologists teach machines to think like exploration geologists—transforming one of humanity’s oldest treasure hunts into a rigorous, interpretable, and data-rich science.

Subject of Research: Machine learning-based mineral prospectivity mapping for orogenic gold deposits in Guangxi, Southwest China

Article Title: A Dual-Driven Mineral Prospectivity Model and Ternary Ore-Control Framework in Guangxi, Southwest China

Article References: Feng, L., Wang, Q., Gregory, D. D., Yang, L., Niu, Y., & Zhao, H. (2026). A Dual-Driven Mineral Prospectivity Model and Ternary Ore-Control Framework in Guangxi, Southwest China. Natural Resources Research. https://doi.org/10.1007/s11053-026-10756-2

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10756-2

Keywords: mineral prospectivity mapping, orogenic gold deposits, machine learning, deep forest, positive-unlabeled learning, Guangxi, Youjiang Basin, Qin-Hang Belt, mineral system knowledge, structural geology, geochemical anomalies, exploration targeting