Ensemble Learning-Based Approach for Email Spam Detection

Abstract

Email spam continues to pose significant security threats, productivity losses, and user inconveniences, making accurate and efficient detection increasingly essential. Traditional machine learning techniques, though widely applied, often struggle with limited accuracy and generalization, necessitating more robust detection strategies. The Enron Email Dataset, sourced from Kaggle, provides a comprehensive collection of email messages, encompassing both spam and legitimate emails. The dataset includes thousands of emails with varied content, formats, and lengths, offering a realistic benchmark for testing spam detection methodologies. The approach employs extensive preprocessing steps, including encoding, data cleaning, tokenization, stopword removal, stemming or lemmatization, TF-IDF vectorization, and oversampling with SMOTE to balance class distribution. Multiple algorithms are applied, including Decision Tree, Random Forest, Naive Bayes, Extra Trees, MLP, SVM, Proposed_Voting, and Proposed_Stacking, with Proposed_Stacking demonstrating superior predictive capability. To further enhance detection performance, Extension_Voting and Extension_Stacking models integrate ensemble strategies to optimize accuracy, precision, recall, and F1-score, achieving the highest result of 98.8%. Along with that, XAI SHAP is utilized as another enhancement for interpretability, and the Flask Framework is employed for the user interface, enabling users to provide inputs and obtain predictions using the trained models. These enhancements suggest promising avenues for more reliable, scalable, and realtime spam detection solutions.