In the literature, a large body of work advocates the use of log-ratio transformation for multivariate statistical analysis of compositional data. In contrast, few studies have looked at how data transformation changes the efficacy of machine learning classifiers within geoscience. This letter presents experiment results and empirical observations to further explore this issue. The objective is to study the effects of data transformation on geozone classification performance when machine learning (ML) classifiers/estimators are trained using geochemical data. The training input consists of exploration hole assay samples obtained from a Pilbara iron-ore deposit in Western Australia, and geozone labels assigned based on stratigraphic units, the absence or presence and type of mineralization. The ML techniques considered are multinomial logistic regression, Gaussian na\"{i}ve Bayes, kNN, linear support vector classifier, RBF-SVM, gradient boosting and extreme GB, random forest (RF) and multi-layer perceptron (MLP). The transformations examined include isometric log-ratio (ILR), center log-ratio (CLR) coupled with principal component analysis (PCA) or independent component analysis (ICA), and a manifold learning approach based on local linear embedding (LLE). The results reveal that different ML classifiers exhibit varying sensitivity to these transformations, with some clearly more advantageous or deleterious than others. Overall, the best performing candidate is ILR which is unsurprising considering the compositional nature of the data. The performance of pairwise log-ratio (PWLR) transformation is better than ILR for ensemble and tree-based learners such as boosting and RF; but worse for MLP, SVM and other classifiers.