Machine learning methods are increasingly applied to mineral prospectivity mapping. However, systematic comparisons between modern ML techniques and traditional methods such as Fuzzy Weights of Evidence (FWoE) remain limited. In this study, we evaluated four machine learning models (Logistic Regression (LR), Support Vector Machine (SVM), Backpropagation Neural Network (BPNN), and XGBoost) alongside the FWoE method for gold prospectivity mapping in a remote region of Labrador, Canada. We used 1,159 lake sediment samples analyzed for 91 geochemical and physical properties. Rather than binary classification, we developed a four class system based on gold concentrations: Background (no gold), Low, Moderate, and High potential.<br />Our results showed that XGBoost achieved the highest macro averaged F1 score (0.279), followed by Logistic Regression (0.270). SVM obtained the highest accuracy (0.724) but this reflects its strong performance on background samples rather than its ability to identify mineralized areas. FWoE scored 0.233 and BPNN scored 0.206. All models performed well on background samples but struggled to distinguish between low, moderate, and high gold classes. Feature importance analysis revealed that geochemical elements including copper, arsenic, and molybdenum were most predictive, though physical properties and field observations also contributed. Our findings indicate that XGBoost is the most effective model for multi class gold mapping, but additional data types such as geological maps and geophysical surveys are needed to improve discrimination between different gold grades.