III: Small: Deep Generative Models for Temporal Graph Generation and Interpretation

III：小：用于时间图生成和解释的深度生成模型

• 批准号: 2007716
• 负责人: Liang Zhao
• 金额: $ 49.81万
• 依托单位: George Mason University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2020-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2007716&HistoricalAwards=false
• 关键词: III; Small; Deep; Generative; Models

Temporal graphs represent a crucial type of data structure where the entities and their connections evolve over time. These time-evolving phenomena are ubiquitous in real-world networks such as social networks, biological networks, and cyber networks. Existing generative models of temporal graphs typically rely on the principles of temporal process of network generation predefined by human heuristics and prior knowledge, such as temporal exponential random graphs, randomized reference models, and dynamic Bayesian models. They usually fit well towards the properties that the predefined principles are tailored for, but usually cannot do well for the others. Unfortunately, the mechanisms of many critical real-world network dynamics are still largely unknown, such as co-evolution of structural and functional connectivities in brain networks, catastrophic cascading failures in power networks, and malware epidemics in the Internet of Things. This project focuses on developing a transformative framework for temporal graphs generative modeling that can automatically learn, characterize, and interpret the underlying patterns and principles from temporal graph observation data. It aims at significantly benefiting the related scientific and engineering domains with open-sourced tools for temporal graphs modeling and network dynamics knowledge distillation. The project includes educational and engagement activities that will substantially increase the community's understanding of temporal graphs.This project will develop a generic framework of generative deep neural networks for temporal graph modeling, generation, and interpretation. The two major types of network dynamic patterns will be investigated, including "topological dynamics of a graph" (e.g., growth of a social network) and "activity dynamics on a graph" (e.g., real-time communications in contact networks). The proposed framework will: 1) automatically learn the (unknown) process of topological dynamics and activity dynamics in discrete- and continuous-time temporal graphs, 2) enforce validity constraints on dynamic topologies and time-evolving activities over the generated temporal graphs, and 3) pursue the temporal graph model interpretability for network dynamic patterns distillation and model intervention. To achieve the above research goals, a number of research activities will be conducted including: i) develop scalable deep generative models for dynamic topologies of large temporal graphs under the temporal-topological constraints, ii) propose novel deep generative models for activity dynamics in temporal graphs under validity-guarantee activation functions, iii) design strategies for modeling the co-evolution of topological dynamics and activity dynamics, iv) pursue model interpretability enhancement by disentangling static and dynamic patterns as well as post-hoc interpretation on the generated temporal graphs by dynamic graph attention and dynamic subgraph detection techniques, and v) develop a novel human-model interaction system for temporal graph visualization and model intervention.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.

时态图代表了一种重要的数据结构类型，其中实体及其连接随时间而演变。这些随时间演化的现象普遍存在于现实世界的网络中，如社交网络、生物网络和计算机网络。现有的时态图生成模型通常依赖于由人类数学和先验知识预定义的网络生成的时态过程的原理，例如时态指数随机图、随机参考模型和动态贝叶斯模型。它们通常很好地适应预定义原则所针对的属性，但通常不能很好地适应其他属性。不幸的是，许多关键的现实世界网络动态的机制在很大程度上仍然是未知的，例如大脑网络中结构和功能连接的共同进化，电力网络中的灾难性级联故障，以及物联网中的恶意软件流行病。该项目的重点是开发一个用于时态图生成建模的变革框架，该框架可以自动学习，表征和解释时态图观察数据的基本模式和原理。它的目的是显着受益于相关的科学和工程领域的时间图建模和网络动力学知识蒸馏的开源工具。该项目包括教育和参与活动，将大大提高社区对时态图的理解。该项目将开发一个生成深度神经网络的通用框架，用于时态图建模，生成和解释。将研究两种主要类型的网络动态模式，包括“图的拓扑动态”（例如，社交网络的增长）和“图上的活动动态”（例如，接触网络中的实时通信）。拟议框架将：1）自动学习离散时间和连续时间时序图中的拓扑动态和活动动态的（未知）过程，2）在生成的时序图上对动态拓扑和时间演进活动实施有效性约束，以及3）追求时序图模型的可解释性以用于网络动态模式提取和模型干预。为达致上述研究目标，我们会进行多项研究活动，包括：i）在时间-拓扑约束下为大型时间图的动态拓扑开发可扩展的深度生成模型，ii）在有效性保证激活函数下为时间图中的活动动态提出新的深度生成模型，iii）用于对拓扑动态和活动动态的协同进化建模的设计策略，iv）通过解开静态和动态模式以及通过动态图注意和动态子图检测技术对所生成的时间图进行事后解释来追求模型可解释性增强，以及V）开发新的人类-该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准。

项目成果

BEAN: Interpretable and Efficient Learning With Biologically-Enhanced Artificial Neuronal Assembly Regularization.

• DOI: 10.3389/fnbot.2021.567482
• 发表时间: 2021
• 期刊: Frontiers in neurorobotics
• 影响因子: 3.1
• 作者: Gao Y;Ascoli GA;Zhao L
• 通讯作者: Zhao L

CPM: A general feature dependency pattern mining framework for contrast multivariate time series

CPM：用于对比多元时间序列的通用特征依赖模式挖掘框架

• DOI: 10.1016/j.patcog.2020.107711
• 发表时间: 2021
• 期刊: Pattern Recognition
• 影响因子: 8
• 作者: Li, Qingzhe;Zhao, Liang;Lee, Yi-Ching;Sassan, Avesta;Lin, Jessica
• 通讯作者: Lin, Jessica

RES: A Robust Framework for Guiding Visual Explanation

• DOI: 10.1145/3534678.3539419
• 发表时间: 2022-06
• 期刊: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining
• 作者: Yuyang Gao;Tong Sun;Guangji Bai;Siyi Gu;S. Hong;Liang Zhao

A Systematic Survey on Deep Generative Models for Graph Generation

• DOI: 10.1109/tpami.2022.3214832
• 发表时间: 2023-05-01
• 期刊: IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
• 影响因子: 23.6
• 作者: Guo, Xiaojie;Zhao, Liang
• 通讯作者: Zhao, Liang

"Temporal Domain Generalization with Drift-Aware Dynamic Neural Networks."

“具有漂移感知动态神经网络的时域泛化。”

• 发表时间: 2023
• 期刊: In The Eleventh International Conference on Learning Representations. 2022.
• 作者: Bai, Guangji;Chen Ling;Liang Zhao.
• 通讯作者: Liang Zhao.