Assessing the Effectiveness of Modern Deep Learning Architectures in Multi-Label News Categorization

Abstract

The rapid spread of fake news online poses a serious threat to public trust, making automated detection systems increasingly necessary. The proliferation of AI-generated content has intensified this challenge, with the number of AI-enabled fake news sites increased tenfold in 2023 alone [1]. This study is the first to compare seven deep learning architectures — CNN, LSTM, Bi-LSTM, RNN, Bi-RNN, GRU, and Bi-GRU — under identical experimental conditions on the same dataset, providing a reliable and unbiased performance benchmark. All models were trained on a publicly available dataset of 44,919 labeled news articles (23,502 fake, 21,417 real) using the Adam optimizer (learning rate = 0.001), binary cross-entropy loss, a batch size of 64, and early stopping, with sequences padded to 500 tokens and an 80/20 train-test split. Each model was evaluated using accuracy, precision, recall, F1-score, and misclassification count. GRU achieved the best overall performance, with 99.91% accuracy, precision, recall, and F1-score, recording only 8 misclassifications on 8,980 test samples. CNN ranked second with 99.82% accuracy and 16 misclassifications, followed by Bi-LSTM at 99.74% with 21 misclassifications. The LSTM achieved 99.57%, while the RNN reached 99.21% Bi-RNN performed the weakest at 96.80% accuracy with 287 misclassifications, the highest error count among all models. Notably, the standard GRU outperformed its bidirectional counterpart, Bi-GRU (99.69%), suggesting that increased architectural complexity does not always yield better performance. These findings demonstrate that GRU is a strong, lightweight choice for real-time fake news detection and offer practical guidance for selecting deep learning models in misinformation filtering systems.