Foundations of Graph Representation Learning in the context of Network and Life Sciences

图神经网络是一种专门用于学习具有图结构的数据的网络。它们在生物学或社交网络等领域的应用非常广泛。尽管它们非常有用，但它们受到图形不随时间变化的假设的影响。考虑到图形动态特性的新框架正在开发中，但它们也没有提供图形动态的全貌。它们通常只考虑图演化过程中少数可能发生的事件，比如插入边缘，但往往忽略节点插入和删除或改变边缘特征等事件。扩展到所有类型的事件将是我的研究重点。完整的深度学习模型忠实地代表了现实世界的数据，可以应用于预测图的未来状态。它也为改进半监督或无监督问题提供了可能性。在这项研究中，我们想研究如何利用关于图的未来状态的知识来提高有关其当前状态的任务的性能。此外，值得注意的是，与动态图的表示学习相关的思想可以应用于该领域之外。因此，改进时态图神经网络的见解可以潜在地应用于其他领域，例如动态图生成。

Graph Neural Networks are specific networks for learning on data with the graph structure. Applications of them arenumerous in fields like biology or social networks. Though extremely useful, they suffer from the assumption that graphsdo not change in time. The new frameworks considering the dynamic nature of graphs are being developed, but theyalso do not offer a full picture of graph's dynamics. They often consider a narrow range of possible events in the graph'sevolution, such as edge insertion, but often ignore events like node insertions and deletions or changing edge features.Extension to all types of events would be my research focus. The complete deep learning model that faithfully representsthe real-world data could be applied to predicting the future states of the graph. It also opens a possibility to improvesolving semi-supervised or unsupervised problems. In this research, we would like to investigate how the knowledge aboutthe future state of the graph can be leveraged to improve performance in tasks concerning its current state. Moreover, itis worth noting that the ideas associated with representation learning on dynamic graphs can be applied outside of thisdomain. Hence insights from improving the temporal Graph Neural Networks could be potentially applied in other areas,such as dynamic graph generation.