Identifying Fake Twitter Trends with Deep Learning

This paper describes a model for detecting and distinguishing fake tweets from real tweets. The model uses one of the most commonly used deep learning algorithms in natural language processing, the Bidirectional Long Short-Term Memory (Bi-LSTM) algorithm. The model was trained on two datasets, the PolitiFact dataset from the FakeNewsNet repository and the PHEME dataset. The datasets are divided into 80% for training and 20% for testing. The model contains two merged models (hybrid model), one for text attributes and the other model for numeric/categorical attributes. To achieve the best result, the model balanced the datasets using four balancing techniques: RandomOverSampler, SMOTE, RandomUnderSampler, and NearMiss. In addition, there are two combination strategies that combine two strategies with certain percentages: combineOverUnderSampling and combineSMOTEUnderSampling balancing strategies. This work is one of the first works to use these combinations and obtain very good results, more than most related studies. For the PHEME dataset, the model was trained with a variety of epochs (10, 20, 50, and 100) to obtain the best result. For the PolitiFact dataset, only 10 epochs were trained at a time. The best result obtained with the PolitiFact dataset is 0.96 for accuracy and precision, 0.95 for recall, F1-score, and ROC-AUC. The best result obtained with the PHEME dataset is 0.87 for accuracy and 0.86 for precision and 0.85 for recall, F1-score, and ROC-AUC.