Tensor Learning Team Seminar (Talk by Mr. Chun WANG from UTS)

-This is an online seminar. Registration is required-
【Tensor Learning Team】
【Date/Time】2021/Aug/6(Fri) 14:00pm-15:00pm JST

【Speaker】 Mr. Chun WANG

Title: Deep Graph Neural Networks for Unsupervised Graph Learning
Abstract:
Graphs are widely used to represent networked data, which contains complex relationships among individuals, and therefore cannot be well represented by traditional flat-table or vector format. Network applications, like social networks or citation networks, have been developing rapidly in recent years. Consequently, graph learning has also attracted much more attention.
Unsupervised graph learning is an important branch of the field since label information is usually not easily accessible. It is much more challenging as unsupervised graph learning aims to model the networked data without training supervision. Associated downstream tasks of unsupervised graph learning may include clustering, link prediction, visualization, etc., which are more challenging without training supervision, and very popular in modern graph analyzing.
We aim to perform effective graph learning, with deep graph neural networks in an unsupervised manner, by solving the following problems: (1) How to efficiently integrate structure and content information of the graph; (2) How to perform end-to-end graph learning for a certain unsupervised task; (3) How to deal with different types of abnormal graph data information.
Specially, we first propose a special marginalized graph autoencoder, to integrate both node content and graph structure information into a unified framework. We add noise to the graph data, and employ a marginalized process for efficient computation; Secondly, we combine graph autoencoder with a self-training model, to conduct a goal-directed training framework. In such a process, clustering and embedding learning are performed simultaneously. Both of them can benefit from the other, thereby learn better graph embedding and clustering. Facing possible data corruption, especially structural corruption for graph data, we develop a dual-autoencoder interaction framework Cross-Graph, which takes advantage of the deep learning memorization effect that DNNs fit clean and easy data first. Two autoencoders filter out untrusted edges alternatively and learn robust embedding from graphs with redundant edges. Finally, to take advantage of possible side information in graph learning, we also propose a contrastive regularized graph autoencoder, that can improve the unsupervised graph learning ability using constraint information. All these frameworks are validated with unsupervised tasks like clustering in the experiments.