Performance Evaluation of ALBERT Model for Fake News Detection Using Kaggle Dataset

Abstract

In the modern digital ecosystem, the widespread dissemination of fake news poses a serious threat to the credibility of online information and social stability. This paper presents a comprehensive evaluation of ALBERT (A Lite BERT), a transformer-based deep learning model, for the detection of fake news articles. Using the Kaggle Fake and Real News Dataset comprising 45,000 labeled news articles evenly divided between fake and real categories, the study assesses ALBERT’s efficiency, accuracy, and reproducibility. The dataset underwent text preprocessing, tokenization, and normalization before fine-tuning the ALBERT-base-v2 model using PyTorch and Hugging Face Transformers. The model achieved an overall accuracy of 94.8%, outperforming traditional machine learning techniques and showing comparable results to larger transformer architectures like BERT and RoBERTa with significantly reduced computational overhead. The findings indicate that ALBERT’s lightweight yet robust architecture enables efficient large-scale deployment in fake news detection systems, providing a balance between computational performance and accuracy.