Comparative Analysis of Machine Learning, Deep Learning, and IndoBERT Models Using SPAMID-PAIR Dataset

Spam comments on social media platforms like Instagram pose a significant challenge, disrupting genuine user interactions. This research investigates and compares the effectiveness of machine learning (ML), deep learning (DL), and IndoBERT models for detecting spam comments on Instagram using the SPAMID-Pair dataset. The central question addressed is how different models perform in classifying spam comments in the Indonesian language, particularly considering contextual relationships between comments and posts. Previous studies have explored various ML and DL models for spam detection, often neglecting the context provided by posts. This research aims to fill this gap by using a dataset that includes both comments and their corresponding posts, highlighting the importance of contextual analysis in improving spam detection accuracy. The methodology involved dataset preparation, preprocessing, data undersampling, and applying various ML, DL, and IndoBERT models. Results indicate that LightGBM, particularly with FastText and BERT embeddings, achieved high accuracy and F1 scores. Deep learning models like BiLSTM, CNN, and Transformer also demonstrated strong performance, with the IndoBERT-large-p1 model outperforming all. This study underscores the significance of advanced NLP techniques and contextual analysis in enhancing spam detection on social media platforms.