Clickbait Detection in Turkish News Headlines Using Machine Learning Methods

Abstract

With the acceleration of access to information in digital media environments, 'clickbait' content presented with misleading and attention-grabbing headlines has become a significant problem that manipulates user perception. This study aims to automatically detect clickbait content in Turkish news texts and, to this end, compares traditional machine learning algorithms (Logistic Regression, Random Forest, XGBoost, CatBoost, LightGBM, SVM) with deep learning-based Transformer models (RoBERTa and BERTurk). In experiments conducted on a dataset consisting of 2,074 manually labeled examples with a balanced class distribution, traditional models were trained using TF-IDF-based vectorization, while Transformer-based models were trained by creating contextual text representations through pre-trained tokenizers. The applied methods were evaluated using classification metrics such as accuracy and F1 score; the highest performance was observed with the RoBERTa model, achieving 91.8% accuracy and 89% F1 score. The results obtained reveal that Transformer-based models with high contextual understanding demonstrate superior performance in clickbait detection compared to traditional methods. Furthermore, the proposed method stands out as a critical tool for digital security, not only in evaluating content accuracy but also in preventing social engineering-based threats such as phishing attacks. © 2025 IEEE.