The Assessment of Machine Learning Model Performance for Predicting Alluvial Deposits Distribution

This paper discusses the development and evaluation of distribution models for predicting alluvial mineral potential mapping. A number of existing models includes Weight of Evidence, Knowledge-driven Fuzzy, Data-driven Fuzzy, Neural-Network, Bayesian Classifier and Geostatistical Kriging. We offer classification models developed in our laboratory, where point pattern analysis was used to identify presence or absence of a known secondary alluvial (cassiterite) deposits in the Nigerian Younger Granite Region (NYGR) and the model performance assessed. We focused on the training and testing data split using longitudinal spatial data splitting (strips and halves) to ensure predictive attribute's independence. The spatial data split runs counter to the traditional random sample data selection as a procedure for checking overfitting of models mainly due to spatial data autocorrelation. Specifically, we used classification algorithms such as; Naive Bayes, Support Vector Machine, K-Nearest Neighbour, Decision Tree Bagging and Discriminant Analysis algorithms for training and testing. We analysed the model's performance results using model predictive accuracy and ROC curve values in two different approaches that improve spatial data independence among predictive attributes to give a meaningful model performance.