Handling Imbalance Noisy Dataset By A Hybrid SMOTE-LOF-Transforms Model

Abstract

In In this study, we aim to evaluate the effectiveness of various machine learning models for fake news detection on PolitiFact and GossipCop datasets obtained from FakeNewsNet, focusing on identifying the most accurate and reliable model. We focused on using state-of-the-art methods combining deep learning transformers and SMOTE resampling techniques for class imbalance, LOF as outlying point detection. The proposed method in this paper, which combines the transformer attention mechanism with SMOTE and LOF resampling techniques, achieved the highest performance metrics on both datasets. The novelty of this model lies in its ability to train on a noisy imbalanced dataset in a short period of time while achieving high accuracy. It recorded an accuracy of 91.5% in PolitiFact and 87.2% in GossipCop, outperforming other models in accuracy, recall and F1 scores. Compared to models such as BERT, BERT + LSTM, and SAFE (multi-faceted), the attention mechanism stands out due to its ability to dynamically focus on relevant items. Overall, this work demonstrates the necessity of retraining models to accommodate the distinctive features of different datasets. It also demonstrates the effectiveness of attention mechanisms in understanding complex narratives and highlights the benefits of a multi-faceted perspective. This indicates that for realistic applications, choosing or designing models with high performance on multiple datasets is necessary, which may pave the way for future work on improving fake news detection using advanced NLP techniques and multimodal approaches.