MemMap: An Adaptive and Latent Memory Structure for Dynamic Graph Learning

Shuo Ji

Abstract: Dynamic graph learning has attracted much attention in recent years due to the fact that most real-world graphs are dynamic and evolutionary. As a result, many dynamic learning methods have been proposed to cope with changes in node states over time. Among these studies, a critical issue is how to update node representations when new temporal events are observed. In this paper, we propose a novel memory structure, Memory Map (MemMap), to address this problem. MemMap is an adaptive and evolutionary latent memory space, where each cell corresponds to an evolving “topic” of the dynamic graph. Moreover, node representations are generated from semantically correlated memory cells rather than from directly linked neighbors. We conduct experiments on real-world datasets and compare our method with state-of-the-art approaches. The results lead to two conclusions: (1) by constructing an adaptive and evolving memory structure during the dynamic learning process, our method effectively captures dynamic graph changes, and the learned MemMap forms a compact evolving structure organized by latent node “topics”; (2) generating node representations from a latent semantic space such as MemMap is more effective and efficient than relying on directly connected neighbors, as in most existing graph learning methods. This is because the number of memory cells in the latent space can be much smaller than the number of nodes in real-world graphs, enabling the representation learning process to balance global and local message passing by leveraging semantic similarity among nodes through correlated memory cells.