Few-Shot Link Prediction for Event-Based Social Networks via Meta-learning.

With the thriving of social network analysis, large efforts have been made on link prediction for event-based social networks (EBSNs). Unfortunately, since society is evolving with constantly emerging social events, it is extremely difficult to accurately capture their semantics and evolution rules at an early stage. Meanwhile, traditional solutions require extensive training from scratch to accommodate new events, leading to lagging predictions and high maintenance costs. To tackle these challenges, we investigate this cross-network few-shot problem and propose a novel meta-learning model for link prediction on new EBSNs. To accurately simulate the few-shot scenarios, we first utilize existing EBSNs to define a task distribution that augments the new event with other observed events. Specifically, we define a unified and generalized target event to be transferred as the few-shot event. Then, we empower a simple but effective event-aware graph attention network to encode existing fine-grained events and the few-shot target events. Furthermore, we follow gradient-based episode learning to obtain transferable knowledge and adapt to unseen EBSNs with sparse connections. Finally, extensive experiments on both public and industrial datasets have demonstrated the performance of fast adaption and even overall performance.