Trade-offs between Accuracy and Efficiency in Fake News Detection

A Comprehensive Study of Lightweight and Deep Learning Techniques

Abstract

The fast spread of fake news on digital platforms needs effective and scalable fake news identification technologies. While deep learning models are very accurate, their computational cost restricts their use in limited-resource scenarios. Using FakeNewsNet dataset, this study provides an evaluation of a lightweight TF-IDF and logistic regression framework against SVM, LSTM, and BERT baseline models across headline-only and full-text scenarios. Experimental results show that the lightweight TF-IDF+LR model achieves competitive accuracy (78.8% on GossipCop and 83.7% on PolitiFact) with dramatic efficiency gains, with 20,000 times faster training than BERT, 0.55 ms inference time, and much smaller memory usage than BERT. An error analysis demonstrates that lexical representations struggle to capture semantic and contextual nuances. It also highlights TF-IDF paired with logistic regression as an acceptable baseline for detecting fake news while outlining its performance limitations. Also, a practical decision matrix is provided for model selection based on the environment's primary constraints.