A Review of Performance of Machine Learning and Deep Learning Models in Multiscale Spam Detection:An Analysis of Feature Engineering Impact Across Email, SMS, and Social Media

Abstract

The increase in spam in email, mobile (SMS), and social media networks requires the constant review of effective detection methods. This paper is a comparative evaluation of the classical Machine Learning (ML) models (Naive Bayes, Logistic Regression, SVM) and the models aimed at advanced Deep Learning (DL LSTM, BERT) and Ensemble Learning models, in terms of their performance in three different communication platforms. The most important element of this study is the analysis of feature engineering methods, such as regular TF-IDF and the use of own vocabulary-based vectorization. Findings have shown that TF-IDF can be highly accurate when using classical models such as Multinomial Naive Bayes (MNB) (98.86% in email), whereas DL models, especially BERT (96.2% in email, 92.9% in SMS, 89.7% in social media), are more generalizable and do not degrade due to spamming strategies present on other platforms. Clear, data-driven recommendations are presented in the model selection based on platform, constraints of computational resources to use, and prioritization of classification metrics in this paper.