Rare-class learning over Mg-doped ZnO nanoparticles

Abstract

This interdisciplinary study is conducted to find answers to two important questions which researchers often face in Machine Learning (ML) and Material Science (MS) fields. In this work, we measure the performance of the most popular ML algorithms (more than a dozen) on rare-class learning problem and determine the best learning algorithm for atom type prediction over the Mg-doped ZnO nanoparticles data obtained from the density-functional tight-binding method. As a result, we observe that tree-based ML algorithms such as Extreme Gradient Boosting (XGB), Decision Trees (DT), Random Forest (RF), outperform other types of ML algorithms, e.g., cost-sensitive learning, prototype models, support vector machines, kernel methods, on both rare-class learning and atom type prediction.