Spam Email Detection Using Advanced Machine Learning Techniques

Abstract

The increasing sophistication of unsolicited and malicious email creates a need for spam-detection methods able to identify context-dependent and linguistically complex patterns that may evade conventional filters. This study compares two approaches to spam classification: Bidirectional Encoder Representations from Transformers (BERT) and Self-Organizing Maps (SOM). Generalization is assessed on five publicly available email benchmarks with different sizes and class distributions. For each dataset, thirty independent train-test partitions are generated; preprocessing, class balancing, model fitting, and REVAC hyperparameter tuning are performed using training data only. Performance on previously unseen test partitions is evaluated using precision, recall, accuracy, and F1-score. Both approaches provide effective classification, but BERT consistently outperforms SOM across all five datasets. The results support the value of contextual, sequence-aware representations for sophisticated spam patterns and underline the importance of rigorous, leakage-free evaluation when comparing spam-detection models.

Keywords: Natural Language Processing (NLP), Transformer Model,  Binary Classification, BERT (Bidirectional Encoder Representations from Transformers), SOM (Self-Organizing Maps).