| Summary: | This thesis explores innovative approaches in graph representation learning and its applications using deep learning models, making significant contributions across several key areas. We first introduce the Graph Meta-Contrast (GMeCo) framework, a novel meta-learning framework for contrastive representation learning on graphs. GMeCo effectively generates augmented graphs and maximizes the mutual information between the augmented graphs and input graphs, outperforming current methods in robust and discriminative feature learning.
Next, we present the Multiresolution Meta-Framelet-based Graph Convolutional Network (MM-FGCN) model. This model represents an advancement in adaptive multiresolution analysis of graphs, overcoming fixed transform limitations and dynamically handling graph data at various scales. MM-FGCN's ability to capture both micro- and macro-level graph structures shows its superiority in various graph learning tasks.
Furthermore, we introduce Graph Spectral Diffusion Model (GSDM), a novel approach for graph-structured data generation. GSDM utilizes low-rank diffusion Stochastic Differential Equations in graph spectrum space, enhancing graph topology generation and reducing computational load. This method demonstrates improved efficiency and quality in graph generation compared to existing models.
Lastly, we develop a novel framework for sequential recommendation systems through a multi-view approach, combining Graph Neural Networks (GNNs) and Transformers. This multi-view structure leverages user-item interactions and collaborative information, offering robust and accurate user preference predictions. This model demonstrates its effectiveness over traditional models.
Overall, this thesis presents effective methods and models in graph representation learning, contributing to advancements in the field and laying a foundation for future research in graph-based deep learning applications.
|
|---|