Self-supervised Probabilistic Graph Structure Learning for Task-agnostic Latent Representation

University Of Texas At Arlington, Arlington TX

Investigators

Abstract

Graphs provide simple and yet powerful mathematical structures to describe pairwise connections among different parties while providing a natural way to develop a deep understanding for real-world environments. There are many situations, however, where graph connections are not readily apparent or are completely hidden. For example, hidden within mountainous microarray data from breast cancer are tree-structure graphs that can delineate breast cancer progressions from one stage to another and thereby are extremely helpful for doctors to devise the best treatment plan for a particular breast cancer survivor. Because they are hidden, the underlying graphical characteristics are not obvious to see and must be learned with intelligent learning models. In this project, the investigators plan to develop and analyze novel graph structure learning models that can uncover latent representations hidden within big data applications. Students will be trained as part of this project, working on the development of mathematical models, numerical algorithms, and software packages for public distribution. This project involves the development and analysis of advanced models and efficient algorithms for latent representation learning via self-supervised graph structure learning. Departing from existing methods, the proposed research tackles task-agnostic graph structure learning so as to not only broaden learning on various types of data, e.g., non-graph data or graph data with unreliable graphs, but also be generalizable, transferrable and robust to different learning tasks. Specifically, new self-supervised probabilistic graph structure learning models, including novel deep graph learning architecture extensions for single and multi-view data, will be formulated to increase the expressive power of the learners, and efficient algorithms to boost task-agnostic graph-based learning from shallow and deep perspectives will be developed to go with the new advanced mathematical models. The research results will appear as a combination of scientific publications and open-source and freely downloadable packages that can be used by researchers in diverse disciplines. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.