Seminar: Graph Representation Learning for Network Generation, Optimization, and Verbalization

_Abstract

Graphs are ubiquitous data structures that denote entities and their relations, such as social networks, citation graphs, and neural networks. The topology of graphs is discrete data, which prevents it from enjoying numerous mathematical and statistical tools that require structured data. Graph representation learning aims to map graphs to their vector representations without substantial information loss, hence paving a new pathway for solving graph problems without discrete algorithms. In this talk, we will first introduce our recent works on graph representation learning that can preserve graphs' geometric information and properties. Then, we will exemplify several interesting research areas where their problem-solving benefits from our leveraging of graph representations.The first area is to solve graph optimization problems, such as influence maximization, source localization, etc., using continuous optimization over graph representations.The second area is to capture and predict deep learning models' dynamics over data distribution drifts, where the graph representation of neural networks is learned to reflect their functional space. The third area is to investigate the correlation and difference of the two views of graphs in mathematical language and natural language, where the graph representation acts as their bridge, with the help of large language models