Clustering in geo-data science: Navigating uncertainty to select the most reliable method

Abstract

Clustering is a fundamental technique in unsupervised learning that groups data points based on their similarities, enabling the discovery of underlying structures and patterns without the need for labeled examples. In geo-data science, clustering plays a pivotal role in applications such as exploratory data analysis, geochemical anomaly detection, and dimensionality reduction for multivariate data analysis. Mineral exploration, in particular, involves various uncertainties that significantly affect decision-making processes. A key source of such uncertainties arises from the choice of clustering methods and their evaluation metrics. Among the widely used clustering methods, K-Means, K-Medoids, Silhouette and Hierarchical Clustering are prominent, with K-Means being a more popular choice in geosciences. Similarly, evaluation techniques such as Silhouette, Davies-Bouldin, Calinski-Harabasz, Elbow, and Bayesian Information Criterion are employed, with the Elbow method being a frequent favorite in geo-data science. However, questions remain regarding the efficiency and suitability of these methods in different contexts. Should we rely solely on K-Means and Elbow, or should we adopt a more case-specific approach, comparing uncertainties and selecting methods that minimize them? This research provides a critical review of clustering methods and evaluation metrics in geo-data science. By developing the pyClusterWise Python package and using illustrative examples, the importance of tailoring these adjustments to the data and selecting clustering techniques and evaluation metrics based on their associated uncertainties is demonstrated. By doing so, the aim is to reduce overall uncertainty and improve decision-making outcomes in mineral exploration.