UMMAN: UNSUPERVISED MULTI-GRAPH MERGE ADVERSARIAL NETWORK FOR DISEASE PREDICTION BASED ON INTESTINAL FLORA

Dingkun Liu, Hongjie Zhou, Yilu Qu, Xiye Niu, Bingchun Luo, Yi Jiang, yongdong xu

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: intestinal flora, graph neural network, unsupervised learning, global embedding

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: Our method combines Graph Neural Network with intestinal flora disease prediction task for the first time and achieves SOTA.

Abstract:

The abundance of intestinal flora is closely related to human diseases, but diseases are not caused by a single gut microbe, but by a combination of a large number of microbial information. It exists a multiplex and implicit connection between gut microbes and hosts, which poses a great challenge to disease prediction through abundance information of gut microbes. Recently, several solution methods have been proposed and have shown the potential of predicting the corresponding diseases. However, these methods have difficulty in learning the inner association between gut microbes and hosts, resulting in unsatisfactory performance. This paper presents a novel architecture, $\textbf{U}$nsupervised $\textbf{M}$ulti-graph $\textbf{M}$erge $\textbf{A}$dversarial $\textbf{N}$etwork (UMMAN). UMMAN can obtain the embeddings of nodes in the Multi-Graph under unsupervised situation, so that it helps learn the multiplex relationship. At the first time, our method combines Graph Neural Network with intestinal flora disease prediction task. We use multiplex relation-types to construct the Original-Graph and destroy the relationship among nodes to get corresponding Shuffled-Graph. We introduce the Node Feature Global Integration (NFGI) module to represent the global features of the graph. Furthermore, we design a joint loss consists of adversarial loss and hybrid attention loss to make real graph embedding agree with the Original-Graph as much as possible and disagree with the Shuffled-Graph as much as possible. Comprehensive
experiments on five classical OTU gut microbiome datasets demonstrate the effectiveness and stability of our method. (We will release our code soon.)

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1172