Deep Representation-aligned Graph Multi-view Clustering for Limited Labeled Multi-modal Health Data

Abstract

Today, many fields are characterised by having extensive quantities of data from a wide range of dissimilar sources and domains. One such field is medicine, in which data contain exhaustive combinations of spatial, temporal, linear, and relational data. Often lacking expert-assessed labels, much of this data would require analysis within the fields of unsupervised or semi-supervised learning. Thus, reasoned by the notion that higher view-counts provide more ways to recognise commonality across views, contrastive multi-view clustering may be utilised to train a model to suppress redundancy and otherwise medically irrelevant information. Yet, standard multi-view clustering approaches do not account for relational graph data. Recent developments aim to solve this by utilising various graph operations including graph-based attention. And within deep-learning graph-based multi-view clustering on a sole view-invariant affinity graph, representation alignment remains unexplored.

We introduce Deep Representation-Aligned Graph Multi-View Clustering (DRAGMVC), a novel attention-based graph multi-view clustering model. Comparing maximal performance, our model surpassed the state-of-the-art in eleven out of twelve metrics on Cora, CiteSeer, and PubMed. The model considers view alignment on a sample-level by employing contrastive loss and relational data through a novel take on graph attention embeddings in which we use a Markov chain prior to increase the receptive field of each layer. For clustering, a graph-induced DDC module is used. GraphSAINT sampling is implemented to control our mini-batch space to capitalise on our Markov prior.

Additionally, we present the MIMIC pleural effusion graph multi-modal dataset, consisting of two modalities registering 3520 chest X-ray images along with two static views registered within a one-day time frame: vital signs and lab tests. These making up the, in total, three views of the dataset. We note a significant improvement in terms of separability, view mixing, and clustering performance comparing DRAGMVC to preceding non-graph multi-view clustering models, suggesting a possible, largely unexplored use case of unsupervised graph multi-view clustering on graph-induced, multi-modal, and complex medical data.