DST: Continual event prediction by decomposing and synergizing the task commonality and specificity

Abstract

Event prediction aims to forecast future events by analyzing the inherent development patterns of historical events. A desirable event prediction system should learn new event knowledge, and adapt to new domains or tasks that arise in real-world application scenarios. However, continuous training can lead to catastrophic forgetting of the model. While existing continuous learning methods can retain characteristic knowledge from previous domains, they ignore potential shared knowledge in subsequent tasks. To tackle these challenges, we propose a novel event prediction method based on graph structural commonality and domain characteristic prompts, which not only avoids forgetting but also facilitates bi-directional knowledge transfer across domains. Specifically, we mitigate model forgetting by designing domain characteristic-oriented prompts in a continuous task stream with frozen the backbone pre-trained model. Building upon this, we further devise a commonality-based adaptive updating algorithm by harnessing a unique structural commonality prompt to inspire implicit common features across domains. Our experimental results on two public benchmark datasets for event prediction demonstrate the effectiveness of our proposed continuous learning event prediction method compared to state-of-the-art baselines. In tests conducted on the IED-Stream, DST’s ET-TA metric significantly improved by 5.6% over the current best baseline model, while the ET-MD metric, which reveals forgetting, decreased by 5.8%.