Class Biased Sarcasm Detection Using Variational LSTM Autoencoder

RumorLLM: A Rumor Large Language Model-Based Fake-News-Detection Data-Augmentation Approach

“…In addition, some researchers have introduced mechanisms such as attention [30] and text-image consistency to achieve better results using text and image information more effectively [31,32]. However, none of the above studies have taken into account the common problem of fake news in the real world-that is, the problem of an unbalanced distribution of fake-news samples in the real world-and the common practice is to resample this category of data from a few samples using three oversampling strategies: Random oversampling; generating synthetic samples from a few categories using k-nearest neighbor methods; oversampling by generating their distribution based on the distribution of the few categories of synthetic samples, oversampling by generating the distribution of the few categories of data samples based on their distributions [33], or even resampling the latent spatial representations mapped by deep learning to balance the dataset by resampling the hidden vectors using a variety of resampling techniques including oversampling, under-sampling and hybrid sampling [11]. However, resampling methods are a single re-use of a particular piece of data, which can introduce noise or cause overfitting problems and do not always improve the performance of the model [34].…”

“…In binary classification, TPR and FPR are defined as follows: TPR = True Positives True Positives + False Negatives (11) FPR = False Positives False Positives + True Negatives (12) ROC-AUC is calculated either by integrating the area under the ROC curve or approximating it using the trapezoidal rule:…”

“…In addition, the larger number of real-news samples can cause the classifier to incorrectly predict new samples as real news when they are actually fake news. This misclassification can lead to significant social harm [5,11].…”