A comparison of tree-based ensemble algorithms on the main element content of monoclinal pyroxene in mafic-ultramafic rocks

Relying on the geochemical composition of the magma tectonic environment to understand the formation process of magma is an important application in rock geochemistry. While the current works to make full use of rock geochemical components for the tectonic setting discrimination are not enough. In this study, the authors utilized four tree-based machine learning methods to make magma tectonic environment discriminations and feature sorting on the 13 main ingredients of monoclinal pyroxene in maficultramafic rocks from global Cenozoic ocean island (OIB), island arc (IAB), and mid-ocean ridge (MORB). Through the comparison of the four tree-based machine learning methods, the authors proved the validity of the tree-based methods for the identification of geochemical components and derived the advantages and disadvantages of the four methods in dealing with the identification of rock tectonic environments:decision trees gain better comprehensibility but have lower recognition accuracy, boosting algorithms AdaBoost and GBDT have the best recognition accuracy but lower comprehensibility, and random forest is a better choice during trading off and comprehensibility performance. Besides, Cr₂O₃, TFeO, TiO₂, FeO and Al₂O₃ are figured out as the most important ingredients for magma tectonic environment discriminations on this dataset.