Modelling particulate matter concentration from loading operations in mineral quarries with a decision tree approach Table 1 Results of normality tests for PM concentration datasets Dataset type raw log(raw) 1/raw ln(raw) raw/1000 PM concentration class TSP PM10 PM2.5 PM1 log(TSP) log(PM10) log(PM2.5) log(PM1) 1/TSP 1/PM10 1/PM2.5 1/PM1 ln(TSP) ln(PM10) ln(PM2.5) ln(PM1) TSP/1000 PM10/1000 PM2.5/1000 PM1/1000 Mean ± SD (µg/m³) 457.91 ± 967.29 308.88 ± 871.79 18.82 ± 39.08 3.52 ± 12.64 2.07 ± 0.74 1.78 ± 0.72 0.89 ± 0.54 0.19 ± 0.49 0.03 ± 0.07 0.05 ± 0.08 0.24 ± 0.28 1.23 ± 3.66 4.76 ± 1.70 4.11 ± 1.66 2.04 ± 1.24 0.44 ± 1.12 0.46 ± 0.97 0.31 ± 0.87 0.02 ± 0.04 0.00 ± 0.01 alphanumeric data and are recognised for their high interpretability and understandability (Gargano, Raggad, 1999; Chien, Chen, 2008; Czajkowski, Kretowski, 2010; Onan, 2015; Aydemir et al., 2020). These qualities make them particularly suitable for simplifying complex data structures and improving decisionmaking processes. Technical advantages such as low cost, fast processing, easy integration into databases, and high reliability have made decision trees among the most preferred classification algorithms (Akçetin, Çelik, 2014; Aydemir et al., 2020). They are widely applied in areas such as knowledge discovery and pattern recognition. Because they produce understandable classification and regression models, decision trees are also extensively used in critical applications such as medical diagnosis and credit risk assessment. Their nonparametric structure enables them to achieve satisfactory accuracy in both classification and regression problems while providing significant interpretability. In regression, decision trees organise independent variables into a hierarchical structure to predict the dependent variable. Starting from the root node, branching is performed through tests on the attributes, and a specific regression equation is presented at each leaf. The estimation process is completed by following the branches based on the results of these attribute tests, starting from the root (Barros et al., 2015; Aksu, 2018). The M5P algorithm is a Java implementation of the M5 algorithm developed by John Ross Quinlan for the WEKA software (Quinlan, 1992; Witten et al., 2011). This algorithm provides a decision tree-based regression method, generating linear regression models at each leaf node. The tree is constructed inductively, with independent variables split into branches based on the dependent variable (Del Campo-Avila, Luengo, 2011; Njeguš, Štavljanin, 2015). The splitting process stops when low variance is observed among the class values at a node or when the number of samples at the node decreases (Öztürk, 2012; Kara, Şamlı, 2021). Pruning is then performed at each leaf node, and linear regression inequalities are derived from the pruned leaves (Sihag, Kumar, 2021). In the The Journal of the Southern African Institute of Mining and Metallurgy Median (min-max) Skewness Kurtosis Sig. 98.40 (1.10 - 6527.90) 46.50 (1.00 - 6527.90) 7.12 (0.30 - 540.22) 1.31 (0.01 - 397.47) 1.99 (0.04 - 3.82) 1.67 (0.00 - 3.82) 0.85 (-0.52 - 2.73) 0.12 (-2.00 - 2.60) 0.01 (0.00 - 0.91) 0.02 (0.00 – 1.00) 0.14 (0.00 - 3.33) 0.76 (0.00 – 100.00) 4.59 (0.10 - 8.78) 3.84 (0.00 - 8.78) 1.96 (-1.20 - 6.29) 0.27 (-4.61 - 5.99) 0.10 (0.00 - 6.53) 0.05 (0.00 - 6.53) 0.01 (0.00 - 0.54) 0.00 (0.00 - 0.40) 3.872 4.862 5.657 18.844 0.104 0.572 0.499 0.370 5.387 4.366 2.583 20.654 0.104 0.572 0.499 0.370 3.872 4.862 5.657 18.813 16.928 25.960 43.898 453.374 -0.303 0.180 0.018 1.276 37.589 26.195 10.783 521.727 -0.303 0.180 0.019 1.276 16.928 25.961 43.902 452.124 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 <0.001 model output, each leaf is presented with two numerical values in parentheses: the first indicates the number of samples in that leaf, and the second shows the ratio of the mean square error of the linear model in the leaf to the global absolute deviation in the entire dataset (Witten et al., 2011; Aydemir, 2018; Altair, 2024; Stackexchange, 2024; WEKA, 2024). This structure combines the explanatory power of a decision tree with the predictive power of linear regression, enabling effective modelling, especially for large and multi-layered datasets. The M5P algorithm provides greater explanatory power and accuracy than classical regression trees, and patterns and correlations in the data can be directly identified through rules and regression equations (Figure 6). The success of numerical models developed using decision tree algorithms such as DecisionStump, RandomForest, RandomTree, and M5P is evaluated using error criteria including the correlation coefficient (R), coefficient of determination (R²), mean absolute error (MAE), root mean square error (RMSE), relative absolute error (RAE), and root relative squared error (RRSE) (Gültepe, 2019). Among these criteria, R² expresses the model's accuracy and ranges from 0 to 1. When this value approaches 0, it indicates that the model does not fit the data (Çınaroğlu, 2016). Similarly, a high correlation coefficient indicates high predictive success; values less than 0.40 are classified as low correlation, 0.40 – 0.70 as normal correlation, and greater than 0.70 as high correlation, indicating better performance (Aydemir, 2018; Schober et al., 2018; Sabti et al., 2019; Tanni et al., 2020). Values close to zero for MAE and RMSE, the most commonly used error measures, indicate high model performance (Wang, Xu, 2004; Gültepe, 2019), and it is recommended to evaluate these two measures together (Chai, Draxler, 2014). If MAE is zero, the model is considered to produce excellent results (Aydemir, 2018), and if RMSE is zero, it indicates ideal performance (Çınaroğlu, 2017). MAE is a fundamental accuracy parameter that measures the average magnitude of errors in the prediction results. It averages the differences between actual and predicted values and indicates how close the predictions are VOLUME 126 MAY 2026 265