Representation learning for temporal networks is challenging, first due to the entanglement of irregular network structures and temporal information, second due to the large-scale essence of real-world temporal networks. Most of previous works for this task have extended graph neural networks to the temporal setting. However, we argue that this type of methods may not well capture some crucial structural features, such as triadic closures, which are crucial for predicting how temporal network evolves over time. In the talk, we will first introduce a strategy to address the problem, named causal anonymous walk that performs online structural feature construction based on sampling walks. Then, we will introduce a very recent work that adopts neighborhood-aware representations to substantially accelerate the above sampling-based approach to make the idea applied to massive networks. The papers to be introduced are:

Inductive representation learning in temporal networks via causal anonymous walks, Wang et al. ICLR 2021

Neighborhood-aware Scalable Temporal Network Representation Learning, Luo & Li, LoG 2022 (selected for oral presentation)