BIGDATA: F: Collaborative Research: Taming Big Networks via Embedding

BIGDATA：F：协作研究：通过嵌入驯服大网络

• 批准号: 1855099
• 负责人: Quanquan Gu
• 金额: $ 49.99万
• 依托单位: University of California-Los Angeles
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-07-01 至 2022-12-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1855099&HistoricalAwards=false
• 关键词: BIGDATA; Collaborative; Research; Taming; Big

In the Internet Age, information entities and objects are interconnected, thereby forming gigantic information networks. Recently, network embedding methods, that create low-dimensional feature representations that preserve the structure of data points in their original space, have been shown to be greatly beneficial for many data mining and machine learning problems over networks. Despite significant research progress, we are still lacking powerful network embedding techniques with theoretical guarantees to effectively deal with massive, heterogeneous, complex and dynamic networks. The PIs aim to develop a new generation of network embedding methods for analyzing massive networks. The research project has the potential to significantly transform graph mining and network analysis. The PIs also plan to develop open course materials and open source software tools that integrate information network analysis and machine learning. This project consists of four synergistic research thrusts. First, it develops model-based network embedding to leverage the first-order and second-order proximity of networks. Second, it devises a family of inductive network embedding methods that are able to leverage both linkage information and side information. Third, it develops both local clustering and deep learning based network embedding methods to attack the complex structure of networks such as locality and non-linearity. Fourth, it develops online and stochastic optimization algorithms for different network embedding methods to tackle the fast growth and evolution of modern massive networks. The new methods developed in this project enjoy faster rates of convergence in optimization, lower computational complexities, and statistical learning guarantees. The targeted applications include but are not limited to semantic search and information retrieval in social/information network analysis, expert finding in bibliographical database, and recommendation systems.

在互联网时代，信息实体和对象相互连接，从而形成巨大的信息网络。最近，网络嵌入方法，创建低维特征表示，保留数据点在其原始空间中的结构，已被证明是非常有益的许多数据挖掘和机器学习问题的网络。尽管研究取得了重大进展，但我们仍然缺乏强大的网络嵌入技术，理论上的保证，以有效地处理大规模，异构，复杂和动态的网络。PI旨在开发新一代网络嵌入方法，用于分析大型网络。该研究项目有可能显著改变图挖掘和网络分析。PI还计划开发集成信息网络分析和机器学习的开放课程材料和开源软件工具。该项目包括四个协同研究重点。首先，它开发了基于模型的网络嵌入，以利用网络的一阶和二阶邻近性。其次，它设计了一个家庭的归纳网络嵌入方法，能够利用链接信息和边信息。第三，它开发了基于局部聚类和深度学习的网络嵌入方法，以攻击网络的复杂结构，如局部性和非线性。第四，它为不同的网络嵌入方法开发了在线和随机优化算法，以应对现代大规模网络的快速增长和演变。该项目开发的新方法在优化中具有更快的收敛速度，更低的计算复杂性和统计学习保证。目标应用包括但不限于社会/信息网络分析中的语义搜索和信息检索、书目数据库中的专家发现以及推荐系统。

项目成果

On the Global Convergence of Training Deep Linear ResNets

• 发表时间: 2020-03
• 期刊: ArXiv
• 作者: Difan Zou;Philip M. Long;Quanquan Gu

Nearly Minimax Optimal Reinforcement Learning for Discounted MDPs

贴现 MDP 的近极小极大最优强化学习

• 发表时间: 2021
• 期刊: Advances in neural information processing systems
• 作者: He, Jiafan;Zhou, Dongruo;Gu, Quanquan
• 通讯作者: Gu, Quanquan

Gradient descent optimizes over-parameterized deep ReLU networks

• DOI: 10.1007/s10994-019-05839-6
• 发表时间: 2019-10-23
• 期刊: MACHINE LEARNING
• 影响因子: 7.5
• 作者: Zou, Difan;Cao, Yuan;Gu, Quanquan
• 通讯作者: Gu, Quanquan

Self-training Converts Weak Learners to Strong Learners in Mixture Models

自我训练将混合模型中的弱学习者转变为强学习者

• 发表时间: 2022
• 期刊: International Conference on Artificial Intelligence and Statistics
• 作者: Frei, Spencer;Zou, Difan;Chen, Zixiang;Gu, Quanquan
• 通讯作者: Gu, Quanquan

Optimization Theory for ReLU Neural Networks Trained with Normalization Layers

• 发表时间: 2020-06
• 作者: Yonatan Dukler;Quanquan Gu;Guido Montúfar