Poster #60 - Taha Mohseni Ahooyi(2)

MOSAIC (Multi-scale Ontology-driven Semantic Analysis and Inference over Contexts): An Optimal Knowledge Graph Embedding Framework

Taha Mohseni Ahooyi(1); Benjamin Stear(1); Dave D. Hill(1); Ryan Corbett(2); J. Alan Simmons(3); Aditya Lahiri(1); Yuanchao Zhang(1); Shiping Zhang(1); James Terry(4); Rebecca Kaufman(1); Asif Chinwalla(1); Christopher Nemarich(1); Adam Resnick(1); Sarah Tasian(1,5); Kristina Cole(1); Jonathan Silverstein(3); Elizabeth Goldmuntz(1,5); Jo Lynne Rokita(2); Sharon Diskin(1,5); Deanne M. Taylor(1,5) (1) The Children's Hospital of Philadelphia, (2) The George Washington University, (3)University of Pittsburg, (4) Drexel University, (5) Perelman School of Medicine, University of Pennsylvania

Knowledge Graph (KG) databases are becoming a mainstay in aggregating ever-increasing complex biological data, providing a unique opportunity to explore and discover seemingly unrelated entities and unforeseen interactions and ushering in a new era in biomedical research. Due to the structural complexity and size barrier associated with most KGs, it is of great practical value to transform the graph structure around biological entities and present each entity as a numerical vector called its node embedding. If done properly, the graph embedding will serve as a computable graph counterpart and can be used as input to a myriad of machine learning applications. In this research we present a novel computational framework to estimate optimal KG embeddings. Our method which is called MOSAIC (Multi-scale Ontology-driven Semantic Analysis and Inference over Contexts) systematically accounts for the local and global structure of the graph surrounding each node and utilizes an iterative optimization algorithm to minimize a distance-based cost function to ensure the computed embeddings preserve the graph structure as closely as possible. The method and its example application will be presented here.