Machine Learning-Based Email Spam Detection: Accuracy, Overfitting and Robustness Analysis

This study evaluates classic and modern machine-learning methods for email spam detection using term frequency and inverse document frequency (TF-IDF) features and a public dataset of ham and spam emails. Nineteen classifiers were trained and compared with accuracy, precision, recall, F1, and variance-based stability. While several models (e.g., Gradient Boosting, Ridge Classifier CV, Bernoulli Naive Bayes) achieved high test accuracy, robustness analysis shows Random Forest and Logistic Regression with cross-validation provide steadier performance and reduced overfitting. Standard-deviation results and train-test gaps expose variance issues in single trees and highlight the practical value of ensembles and regularized linear models. The work underscores that deployment choices should favor consistent, generalizable behavior over peak scores alone.