Generalizing Graph Diffusion Models
Event details
| Date | 02.02.2024 |
| Hour | 13:00 › 15:00 |
| Speaker | Yiming Qin |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Maria Brbic
Thesis advisor: Prof. Pascal Frossard
Co-examiner: Prof. Olga Fink
Abstract
Generating new graph structures is crucial for a wide range of tasks such as drug discovery and protein design.
Particularly, diffusion-based models have set benchmarks in the domain of graph generation.
Building upon this success, there is a growing research interest in exploring how graph diffusion models can be generalized across different contexts.
In this report, we first review foundational work in graph diffusion, detailing diffusion models and elaborating on how they are effectively applied in the graph domain.
Then, due to the square space complexity of such models, we proceed to introduce a new approach that generalizes current models to generate larger graphs and maintains high expressivity through multiscale generation.
Afterward, we explore a study that interprets diffusion models from an optimal transport perspective, providing an algorithm that extends these models to tackle more complex distribution shift tasks. This advancement facilitates a range of potential downstream applications
Finally, we discuss future research directions in graph generation, with a particular emphasis on enhancing the generalization capacity of current models.
Background papers
Exam president: Prof. Maria Brbic
Thesis advisor: Prof. Pascal Frossard
Co-examiner: Prof. Olga Fink
Abstract
Generating new graph structures is crucial for a wide range of tasks such as drug discovery and protein design.
Particularly, diffusion-based models have set benchmarks in the domain of graph generation.
Building upon this success, there is a growing research interest in exploring how graph diffusion models can be generalized across different contexts.
In this report, we first review foundational work in graph diffusion, detailing diffusion models and elaborating on how they are effectively applied in the graph domain.
Then, due to the square space complexity of such models, we proceed to introduce a new approach that generalizes current models to generate larger graphs and maintains high expressivity through multiscale generation.
Afterward, we explore a study that interprets diffusion models from an optimal transport perspective, providing an algorithm that extends these models to tackle more complex distribution shift tasks. This advancement facilitates a range of potential downstream applications
Finally, we discuss future research directions in graph generation, with a particular emphasis on enhancing the generalization capacity of current models.
Background papers
- Permutation Invariant Graph Generation via Score-Based Generative Modeling (Chenhao Niu, Yang Song, Jiaming Song, Shengjia Zhao, Aditya Grover, Stefano Ermon), https://arxiv.org/abs/2003.00638
- Efficient and Scalable Graph Generation through Iterative Local Expansion (Andreas Bergmeister, Karolis Martinkus, Nathanaël Perraudin, Roger Wattenhofer), https://arxiv.org/abs/2312.11529
- Diffusion Schrödinger Bridge with Applications to Score-Based Generative Modeling (Valentin De Bortoli, James Thornton, Jeremy Heng, Arnaud Doucet), https://arxiv.org/abs/2106.01357
Practical information
- General public
- Free