580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can spatial and spectral domains be unified to provide a comprehensive framework for Graph Neural Networks (GNNs)?
Graph Neural Networks (GNNs) have been developed through different theoretical lenses, primarily in the spatial and spectral domains, each proposing distinct models and mechanisms for graph representation learning. This segregation complicates model selection, understanding, and comparison, hindering theoretical insights and practical deployment. Bridging this divide seeks to develop a coherent, unified framework that reconciles the spatial and spectral approaches, elucidates their interrelationships, and provides rigorous theoretical grounding for diverse GNN models. This unification is critical for demystifying GNN effectiveness, improving model interpretability, and fostering systematic developments in graph representation learning.
Key finding: This study proposes a unifying theoretical framework that formally connects spatial and spectral GNN models through spectral graph theory and approximation theory, establishing an equivalence between spatial connection... Read more
Bridging the Gap between Spatial and Spectral Domains: A Unified Framework for Graph Neural Networks
Key finding: This work closely complements and extends prior efforts by synthesizing spectral graph theory with approximation insights to build a unified framework capturing both spatially and spectrally motivated GNN models. It... Read more
Key finding: This monograph section reviews graph topology learning, highlighting the challenge of unknown graph structures in data analysis. It elucidates how integrating graph topology inference with graph neural network methodologies... Read more
2. What are effective mathematical and algorithmic models for representing, analyzing, and visualizing dynamic and complex graph structures?
As graph-structured data grow in size and complexity across application domains, mathematical models and tools that capture their dynamic behaviors and multi-layered nature are essential. This research theme focuses on developing rigorous representations—such as multi-level, multiplex, and temporal graph models—and efficient algorithmic tools for graph analysis and visualization. It emphasizes the challenge of representing networks with layered or evolving connectivity patterns (e.g., networks of networks), the transformation of graphs into signal representations to facilitate analysis, and the design of scalable visualization solutions that reveal algorithmic dynamics and graph structural properties over time. Effective solutions in this area are critical for enabling insights into complex systems in biology, social science, transport, and computer science.
Key finding: This work formalizes multilevel networks where multiple networks defined on the same node set are interconnected through a higher-level network on the layers themselves. By representing ties both among nodes within layers and... Read more
Key finding: This paper proposes a novel method to analyze dynamic graphs by transforming each time-step graph into a collection of signals via multidimensional scaling, facilitating spectral analysis of evolving graph structures. The... Read more
Key finding: The authors propose an interactive web-based visualization tool that combines time-to-space static mappings and animated time-to-time mappings to analyze and explore the step-wise execution dynamics of graph algorithms over... Read more
Key finding: Graph Investigator is introduced as a versatile software tool enabling comprehensive complex network analysis by extracting over eighty statistical and algebraic network descriptors. It facilitates global and local structural... Read more
3. How can signal processing frameworks be adapted for effective representation, sampling, and learning of data on graphs?
Graph Signal Processing (GSP) extends classical signal processing to data defined on irregular graph domains, enabling novel approaches for graph signal representation, filtering, sampling, and learning. This theme investigates explicit graph signal models—such as smooth, piecewise-constant, and piecewise-smooth signals—each associated with tailored dictionary constructions that promote sparsity and effective approximation. It also studies how these frameworks facilitate graph signal recovery, compression, and change detection, crucial for tasks like sensor network data analysis and environmental monitoring. With rising integration into machine learning pipelines, these representations underpin advanced analysis and inference over graph-structured data.
Key finding: This foundational work formalizes a unified representation-based framework for graph signals, categorizing signals into smooth, piecewise-constant, and piecewise-smooth classes. For each class, dedicated graph dictionaries... Read more
Key finding: Building on fundamental graph spectral analysis, this monograph segment establishes core signal processing concepts on graphs, including graph signal shifts, graph Fourier transforms, and graph filtering. It introduces... Read more