Meta-Learner-Based Frameworks for Interpretable Email Spam Detection

Meghana Kshirsagar^1*Vedant Rathi²Conor Ryan^1*

• ¹University of Limerick, Limerick, Ireland
• ²University of Illinois at Urbana-Champaign, Champaign, Illinois, United States

Introduction: With the increasing reliance on digital communication, email has become an essential tool for personal and professional correspondence. However, despite its numerous benefits, digital communication faces significant challenges, particularly the prevalence of spam emails. Effective spam email classification systems are crucial to mitigate these issues by automatically identifying and filtering out unwanted messages, enhancing the efficiency of email communication. Methods: We compare five traditional machine-learning and five deep-learning spam classifiers against a novel meta-learner, evaluating how different word embeddings, vectorization schemes, and model architectures affect performance on the Enron-Spam and TREC 2007 datasets. The primary aim is to show how the meta-learner's combined predictions stack up against individual ML and DL approaches. Results: Our meta-learner outperforms all state-of-the-art models, achieving an accuracy of 0.9905 and an AUC score of 0.9991 on a hybrid dataset that combines Enron-Spam and TREC 2007. To the best of our knowledge, our model also surpasses the only other meta-learning-based spam detection model reported in recent literature, with higher accuracy, better generalization from a significantly larger dataset, and lower computational complexity. We also evaluated our meta-learner in a zero-shot setting on an unseen real-world dataset, achieving a spam sensitivity rate of 0.8970 and an AUC score of 0.7605. Discussion: These results demonstrate that meta-learning can yield more robust, bias-resistant spam filters suited for real-world deployment. By combining complementary model strengths, the meta-learner also offers improved resilience against evolving spam tactics.