現實世界中許多高維度不規則區域的資料可以用圖來表示，像是社群網路、大腦連接組、詞向量。有多項研究是在探討如何以圖的架構開發模型，如：Graph Neuron Network、Graph Convolutional Network和Graph Attention Network等，這些網路是在靜態圖上運作的，多數是用於分類任務。 在這項研究中，我們著重於動態序列資料的預測，像是每年co-author relation graph的變化，或是玩家在隨時間的互動關係。我們的模型結合了Variational Graph Auto-Encoder(VGAE)、Graph Convolutional Network(GCN)和Long short-term memory(LSTM)，其中廣泛的使用遞迴神經網路。我們在五個資料集上檢測模型的效果，High-energy physics theory citation network、Dynamic Face-to-Face Interaction Networks、CollegeMsg temporal network、Email-Eu-core temporal network和DBLP collaboration network and ground-truth communities。;A lot of real-world data in high-dimensional irregular domain, such as social networks, brain connectomes or words’ embedding, can be represented by graphs. There are various researches that developed models to operate on graph-structured data, such as Graph Neuron Network, Graph Convolutional Network, Graph Attention Network, etc. These networks are operated on static graphs, mostly use for classification task. In this research, we are interested in prediction of dynamic and sequential data, such as how is the changes of the co-author relationship graph in each year, or how players interact with each other through time. We implement a model from the combination of Variational Graph Auto-Encoder (VGAE), which is a variant of Graph Convolutional Network (GCN) with Long short-term memory (LSTM), which is a wildly used Recurrent Neural Network. And evaluate the performance of the model on 5 datasets, High-energy physics theory citation network, Dynamic Face-to-Face Interaction Networks, CollegeMsg temporal network, Email-Eu-core temporal network and DBLP collaboration network and ground-truth communities.
