PhD Dissertation Proposal by LEE Ween Jiann | Incorporating Intrinsic Structures into Entity Matching and Representation Learning

With participation on the internet becoming a norm, we can harness the exponential growth of user-generated content to improve the accuracy of deep learning models. Content is presented in a wide variety of formats such as text, visual and spatial which when facilitated by multi-modality learning would help us to gain deeper insights than relying on a single modality alone. However, often, this data is largely unstructured and scattered over various locations. The aim is to explore ways to merge and unify these into a single location and encode it into a structured form which we can easily use in creating better models such as for use in recommendation systems. In this talk, we present our solutions to address this problem.

Inspired by a fundamental problem to match entities across two data sources, and the presumption that entities to be matched are of comparable granularity, we present a new approach that abolishes this presumption and addresses the scenario where entities have varying granularity. Additionally, we address the spatial aspect of material entities that are located physically in the world. As we are constrained by space-time, the physical locality of a material entity is an important aspect that prior graph neural network approaches to representation learning have fallen short to consider. With this, we present two new representation learning models that are able to learn spatially-aware representations for geospatial entities. Finally, we look toward the future and present our plans to merge, unify and encode data to improve tomorrow's model.

Ween Jiann Lee is a Ph.D. candidate at the School of Computing and Information Systems, Singapore Management University (SMU), and working toward a Ph.D. degree in computer science. He received a bachelor’s degree in business management majoring in Operations Management and Analytics from SMU. His research interest lies in graph neural network models, representation learning, and matching problems.