Continual Learning for Fake News Detection from Social Media

The prevalence of fake news over social media has a profound impact on justice, public trust and society as a whole. Although significant effort has been applied to mitigate its negative impact, our study shows that existing fake news detection algorithms may perform poorly on new data. In other words, the performance of a model trained on one dataset degrades on another and potentially vastly different dataset. Considering that in practice a deployed fake news detection system is likely to observe unseen data, it is crucial to solve this problem without re-training the model on the entire data from scratch, which would become prohibitively expensive as the data volumes grow. An intuitive solution is to further train the model on the new dataset, but our results show that this direct incremental training approach does not work, as the model only performs well on the latest dataset it is trained on, which is similar to the problem of catastrophic forgetting in the field of continual learning. Instead, in this work, (1) we first demonstrate that with only minor computational overhead, balanced performance can be restored on both existing and new datasets, by utilising Gradient Episodic Memory (GEM) and Elastic Weight Consolidation (EWC)-two techniques from continual learning. (2) We improve the algorithm of GEM so that the drop in model performance on the previous task can be further minimised. Specifically, we investigate different techniques to optimise the sampling process for GEM, as an improvement over random selection as originally designed. (3) We conduct extensive experiments on two datasets with thousands of labelled news items to verify our results.

Automated summary

The prevalence of fake news on social media has a significant impact on justice, public trust, and society. Existing fake news detection algorithms may perform poorly on new data, necessitating the use of continual learning techniques to maintain model performance across different datasets. This study demonstrates that with minor computational overhead and optimization of techniques like GEM, balanced performance can be achieved on both existing and new datasets.