Protein methylation is vital in post-translational regulation, making accurate site prediction crucial for biomedical research and drug development. A key challenge is class imbalance in datasets. This study addresses the class imbalance problem using K-fold stratified cross validation and SMOTE and compares their performance on five machine learning (ML) algorithms. Using validated methylated and unmethylated Homo sapiens protein sequences, we extracted sequence-based features such as physicochemical properties, one-hot encoding, Shannon entropy, k-mer composition, and k-spaced amino acid pair frequency, forming a 1266-dimensional feature vector. We applied ML algorithms to three methyl-arginine datasets: mono-methyl, asymmetric di-methyl, and symmetric di-methyl arginine. SMOTE-based class balancing achieved 91.6% accuracy with an F1-score of 0.91. Our results highlight the effectiveness of SMOTE combined with ML in predicting arginine methylation sites.