What

Learning structural and temporal relationships on dynamic graphs with node-level and edge-level changes taking place. Best to understand this paper from a "social media graph" perspective.

The paper provides a framework for general temporal graph networks that most other

Why

• Current graph methods mostly work on static graphs but most real-life situations are dynamic
  • This paper offers a solution for continuous time dynamic graphs (CTDG) as opposed to other methods that work on discrete time dynamic graphs (DTDG) where events take place over equal time intervals.

How

There are two types of events that happen in a graph

• Node-level events → events that take place within a node $i$ in isolation (internal change)
• Intersection events → events that take place between two nodes $i$ and $j$ (edge event)

<aside> ⭐ Deletion events are not the focus of this paper but are explained in the appendix.

</aside>

The paper presents a modular approach to learning structural-temporal features of graphs:

• Memory
• Message Passing Function
• Message Aggregator
• Memory Updater
• Embedding

Training

• Has a Raw Message Store that keeps track of messages from previous batches

Datasets