A systematic evaluation of data preprocessing and model optimization for machine learning algorithms: Using sphalerite trace element data as an example

Abstract

Mineral chemistry combined with machine learning (ML) algorithms has been widely used in ore genesis and mineral exploration. However, a systematic investigation on the effect of data preprocessing and optimal parameters on the performance of ML methods was not enough. Based on a real unbalanced dataset with 4,312 trace element data of sphalerite from different types of deposits, this study investigated the classification performance of eight ML algorithms, including unsupervised principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), and supervised linear discriminant analysis (LDA), partial least squares-discriminant analysis (PLS-DA), random forest (RF), support vector machine (SVM), and extremely greedy tree boosting (XGBoost). Different preprocessing methods, including missing value imputation, data transformation, and feature selection, were evaluated. For skewed distribution data, the study recommends using KNN, Mode, and Median imputation methods to reduce bias. The PCA, t-SNE, LDA, PLS-DA, and SVM algorithms perform well on data transformed by CLR or logarithmic methods, while RF and XGBoost algorithms also show good classification performance on untransformed data, especially XGBoost shows the best performance on data without imputation. Using the Recursive Feature Elimination (RFE) feature selection with the feature importance, the key features with the most discriminative capabilities can be screened out. This research not only optimizes mineral data processing workflows but also provides important support for improving the precision and accuracy of ML models.