Comparative Performance of Naïve Bayes and Support Vector Machine Algorithms for Machine Learning-Based Spam Email Detection

Email spam is one of the primary threats in digital communication with continuously growing volume. Based on the Spam and Phishing in 2024 report released by Kaspersky, email spam accounted for approximately 47.27% of total global email traffic in 2024. This study aims to implement and compare the performance of the Naïve Bayes and Support Vector Machine (SVM) algorithms in detecting email spam using the Enron Spam Subset dataset from Kaggle consisting of 10,000 emails with a balanced distribution (5,000 spam, 5,000 ham). The research stages include preprocessing text (case folding, cleaning, tokenizing, stopword removal, stemming), TF-IDF feature extraction with N-gram (unigram and bigram), and evaluation using 5-Fold Stratified Cross Validation and Paired Sample t-Test (α = 0.05). The results show that both NB+Unigram and SVM+Unigram configurations achieved the highest accuracy (98.80%) on the test set, with NB+Unigram achieving an F1-score of 0.9880, while SVM+Unigram achieved an F1-score of 0.9881. In the cross validation evaluation, SVM+Unigram obtained the highest mean F1-score of 0.9827. Statistical testing confirmed a significant difference between NB and SVM in the unigram configuration (p = 0.023). The TF-IDF unigram combination yielded higher performance than bigram for both algorithms in the context of this dataset.