Learning temporal attention in dynamic graphs with bilinear interactions

Reasoning about graphs evolving over time is a challenging concept in many domains, such as bioinformatics, physics, and social networks. We consider common case which edges can be short term interactions (e.g., messaging) or long structural connections friendship). In practice, are often specified by humans. Human-specified both expensive to produce suboptimal for the downstream task. To alleviate these issues, we propose model based on temporal point processes variational autoencoders that learns infer attention between nodes observing node communication. As drives between-node feature propagation, using dynamics of learn this key component provides more flexibility while simultaneously avoiding issues associated with human-specified edges. also bilinear transformation layer pairs features instead concatenation, typically used prior work, demonstrate its superior performance all cases. experiments two datasets dynamic link prediction task, our outperforms baseline requires graph. Moreover, learned semantically interpretable infers similar actual graphs.