基于隐特征分析的信息推荐技术研究

Latent factor analysis is an important branch of information recommendation techniques. It focuses on building latent factor models on high-dimensional, sparse and fractional data. Current works mostly focus on single characteristic of latent factor models and single data analysis task. Systematic research regarding various characteristics of latent factor models is rarely seen. This project focuses on latent factor analysis-based information recommendation, and conducts the research from the aspects of model generalization, data representation and temporal migration. First, we will focus on the highly-efficient first-order and second-order solvers for latent factors, regularization schemes for latent factor models on high-dimensional and sparse matrices, hierarchical latent factor models combining the deep learning principle, and ensemble methods for latent factor models for enhancing the generalization of latent factor models. Second, we will investigate the non-negative latent factor models, symmetric latent factor models and multi-objective latent factor models for improving the representativeness of latent factor models on constrained data. Third, we will build entity-entity-temporal tensors based on high-dimensional and sparse matrices with timestamps, and then build latent factor models on tensors, making them correctly describe the temporal patterns to possess the ability of temporal migration. The outcome of this project will promote the research progress of information recommendation techniques, as well as serve industrial and big-data related applications.

隐特征分析是信息推荐技术的重要分支，重点是在高维、稀疏、碎片化的数据上构建隐特征分析模型。现有工作大多只关心隐特征模型的单一方面特性和单一数据分析任务，缺乏系统研究。本项目聚焦基于隐特征分析的信息推荐问题，从模型泛化能力、数据表征能力和时间迁移能力三个角度开展研究。首先，在隐特征一阶和二阶高效求解方法、高维稀疏矩阵上隐特征分析的正则化方法、结合深度学习原理的多层隐特征分析模型和适用于隐特征分析模型的集成学习方法四个方面开展研究，以强化模型泛化能力；其次，研究非负隐特征分析模型、对称隐特征分析模型和多目标隐特征分析模型，以提升模型对带约束数据的表征能力；最后，研究如何基于带时间标记的高维稀疏矩阵构建实体-实体-时间张量，并研究在实体-实体-时间张量上的隐特征分析模型，使其能通过正确描述时变模式以具备时间迁移能力。相关研究成果既能推动信息推荐研究进展，又能应用于实际大数据系统。

隐特征分析是信息推荐技术的重要分支，重点是在高维、稀疏、碎片化的数据上构建隐特征分析模型。现有工作大多只关心隐特征模型的单一方面特性和单一数据分析任务，缺乏系统研究。本项目聚焦基于隐特征分析的信息推荐问题，从模型泛化能力、数据表征能力、时间迁移能力和高效精确优化四个角度开展研究，相关研究成果既能推动信息推荐研究进展，又能应用于实际大数据系统。基于上述研究内容，超额完成研究任务，发表论文40篇（其中包括33篇IEEE汇刊论文）、申请国家发明专利10项、培养人才14人、获得省部级一等奖1项、本人入选国家级人才计划。