Perbandingan Machine Learning dan Deep Learning untuk Klasifikasi Spam dengan Class Balancing

Abstract

SPAM/HAM classification faces two primary challenges: class imbalance and the need to select a model that is both accurate and efficient. This study compares The Conventional Machine Learning Algorithm Naive Bayes and Support Vector Machine with Recurrent Deep Learning Algorithm Long Short-Term Memory, and Gated Recurrent Unit using a class-balancing strategy. The data underwent cleaning and deduplication before being split into a development set (80%) and a final test set (20%). Balancing was applied exclusively to the training set by downsampling HAM and oversampling SPAM to reach 1,500 samples each, while the validation and final test sets retained their natural distributions. NB and SVM utilized TF-IDF, whereas LSTM and GRU employed tokenization, 128-dimensional embeddings, and padding masks. NB achieved an accuracy of 98.63% and a Macro-F1 score of 96.65%, the highest among the four models. SVM yielded a PR-AUC of 97.55%; GRU achieved the highest SPAM recall at 94.96%; and LSTM attained the highest ROC-AUC at 98.92%. NB's training time was merely 0.004 seconds, compared to 6.61 seconds for GRU and 12.95 seconds for LSTM. The results demonstrate that traditional models remain competitive, offering a superior performance-efficiency trade-off for short-message data following class balancing.