Health and Behavior: DEEP LEARNING FOR UNSTRUCTURED DATA ANALYTICS AND MINING

REU Scholar: Richard Gao

REU Scholar Home Institution: Rice University

REU Mentor: Xingquan Zhu, Ph.D.

Project: Graph Learning For Network Data

Many datasets naturally lend themselves to be structured as graphs. Leveraging their topological information allows Graph Neural Networks (GNN) to outperform their non-graph based counterparts. However, most real world graph datasets are heterogeneous and cannot be processed by traditional GNNs due to inconsistent node and edge typing. In this presentation, we introduce a homogenizing pipeline that will allow us to modify these data sets to be compatible with current GNN technology. By choosing a specific metapath and nodetype, we can transform heterogeneous, single-labeled graph data into homogeneous, multi-labeled data that we can pass through traditional GNN structures. The multi-labeled nature of the transformed data requires us to adapt traditional, single-label classification models to scale to the multiple-labels, or create new, multi-label specific models. Then, by evaluating these different GNN architectures on the transformed heterogeneous graphs and comparing their performances, we can find the optimal architecture for the multi-label node classification task in the context of heterogeneous graphs.