Machine Learning-Based Spam Email Classification

Abstract

Electronic communication forms the foundation of today's digital infrastructure. Contemporary filter systems need to handle high-dimensional textual sparsity and fluctuating vocabulary distributions continuously by bypassing static defenses. Fluctuations in the dataset and an increase in the demand for low-latency inference at some network edge nodes have also raised the operating risk of the enterprise infrastructure. Given the above problems, the low-complexity operational baseline is provided by traditional probabilistic frameworks. Deep neural networks and pre-trained large language models (LLMs) can achieve contextual zero-shot classification but are computationally intensive and lack interpretability. Tree-based ensemble classifiers have achieved a good trade-off between prediction accuracy and inference speed for high-dimensional text matrices. Random forest models use parallel dual-randomness sampling and majority voting to handle token noise, and XGBoost employs second-order Taylor approximations and L2 regularization for fast convergence under strict latency constraints and class imbalance. This paper provides a mathematical foundation for well-known tree-based ensemble classifiers, including XGBoost and random forest. By investigating the basic structure of features and optimisation surfaces, this paper proposes all-encompassing structural plans for building high-adaptability text classification systems that can handle various data shifts.