EvolveDetector: Towards an evolving fake news detector for emerging events with continual knowledge accumulation and transfer

The prevalence of fake news on social media poses devastating and wide-ranging threats to political beliefs, economic activities, and public health. Due to the continuous emergence of news events on social media, the corresponding data distribution keeps changing, which places high demands on the generalizabilities of automatic detection methods. Current cross-event fake news detection methods often enhance generalization by training models on a broad range of events. However, they require storing historical training data and retraining the model from scratch when new events occur, resulting in substantial storage and computational costs. This limitation makes it challenging to meet the requirements of continual fake news detection on social media. Inspired by human abilities to consolidate learning from earlier tasks and transfer knowledge to new tasks, in this paper, we propose a fake news detection method based on parameter-level historical event knowledge transfer, namely EvolveDetector, which does not require storing historical event data to retrain the model from scratch. Specifically, we design the hard attention-based knowledge storing mechanism to efficiently store the knowledge of learned events, which mainly consists of a knowledge memory and corresponding event masks. Whenever a new event needs to be detected for fake news, EvolveDetector retrieves the neuron parameters of all similar historical events from the knowledge memory to guide the learning in the new event. Afterward, the multi-head self-attention is used to integrate the feature outputs corresponding to these similar events to train a classifier for the new event. Experiments on public datasets collected from Twitter and Sina Weibo demonstrate that our EvolveDetector outperforms state-of-the-art baselines, which can be utilized for cross-event fake news detection.