Custom Loss Functions to Tackle Class Imbalance, Enhance Specific Performance Measure, And Improve Classification Performance in The Context of Fake News Detection

Abstract

The spread of false information on social media has raised serious concerns across the globe. Most social media contents are multimodal, comprising text, images, and substantial data generated through likes, tweets, retweets, and replies. To address this challenge, a hierarchical deep learning model that learns and fuses features from each modality is necessary. In the literature, early and decision level fusion inspired hierarchical deep learning multimodal models are designed for fake news detection. However, these models rely solely on cross-entropy loss, which is insufficient given the real-world, imbalanced nature of fake news datasets. Additionally, decision-level fusion models present simplified architectures but they cannot effectively handle the complexities of multimodal data resulting in less accurate models. In this work, we address the mentioned issues by introducing custom loss functions specific to early and decision level fusion multimodal deep learning models. Three different loss functions namely: Adaptive Hybrid Loss (AHL), False Negative based loss (FNL) and Similarity preserving loss (SPL) are designed. The proposed loss functions combine the benefits of focal loss and cross-entropy loss. AHL is suitable for the problems where achieving precision and recall are equally important, FNL is used in case where recall is more important as in case of Fake news detection. SPL provides balance between precision and accuracy and improves the detection accuracy. On Fake news net dataset AHL provides a F1 score of 92.36% and FNL provides improved recall of 93.18% and SPL provides F1 score of 90.26% which is higher than the models trained using cross entropy loss. Our experimental results demonstrate that these tailored loss functions significantly boost the performance of multimodal fake news detection(FND) systems, providing a comprehensive solution for combating misinformation in the digital age.