面向金融大数据的半监督聚类集成挖掘关键技术研究

The financial big data mining is a challenging research project with clear application prospect in fields like financial risk control, financial management and decision optimization, business innovation. Financial data are massive, high-dimensional and nonlinear, whereas available data mining methods are usually complex computation and with weak adaptability. Some researches on key techniques based on financial data clustering ensemble mining are carried out. The investigations were based on semi-supervised theory. The main contents including: (1) some strategies such as weight operator, distance measurement and multi-group intelligent evolution were developed to propose a self-adaptive intelligent weighted similarity measuring method. (2) Based on singular value decomposition and stable threshold, novel semi-supervised clustering models combined with multiple strategies are established. (3) The diversity clustering factor and measure method are both proposed, so the design of ensemble function in ensemble model is also obtained. (4)The novel models and technologies are established based on semi-supervised theory, clustering methods and ensemble technologies, which have some advantages such as robustness, accuracy, parallelism. In addition, the mathematical proof of convergence is also carried out. The objective of this project is to solve the core problems influencing the data mining performance in financial field, and provide theoretical and technical supports for financial data mining, intelligent information processing,financial management and decision-making.

金融大数据挖掘是一个具有挑战性的研究课题，在金融风险防控、金融管理与决策优化、业务创新等方面具有明确的应用前景。针对金融数据海量、高维、非线性等特点和现有挖掘方法存在的计算复杂度高、普适性差等问题，本项目拟结合半监督学习理论，开展金融数据聚类集成挖掘的关键技术研究。主要内容包括：（1）分别构造赋权算子、距离测度和多群体智能进化等策略，提出自适应智能赋权相似度度量方法。（2）分别构造奇异值分解、稳定阈值等策略，提出融合多种策略的半监督聚类新模型。(3)构建差异性基聚类因子和差异性度量方法，实现集成模型中共识函数的设计。（4）在此基础上，融合半监督理论、聚类方法和集成技术，构建具有鲁棒性、精确性和并行性的半监督聚类集成新模型与新技术，并对其收敛性进行数学证明。本项目拟解决影响金融数据挖掘性能的核心问题，可望为金融数据挖掘、智能信息处理、金融管理与决策提供理论与技术支持。

金融数据具有海量、高维和非线性等特点。金融数据挖掘作为金融数据分析与处理的重要技术，目前已经成为国内外研究的热点，并成为独具特色的研究分支。聚类是数据挖掘的关键技术之一，是本项目研究的重要手段。本项目针对传统聚类算法密度峰值聚类算法、仿射传播聚类算法等进行了理论改进研究与应用研究。考虑到传统聚类算法存在的弊端，通过引入布谷鸟搜索算法、蝙蝠优化算法、萤火虫优化算法等若干具有快速全局搜索优势的群智能优化方法，提出了若干具有仿生智能特征的新的聚类方法，如基于蝙蝠优化的再聚类算法，基于萤火虫优化的近邻传播聚类算法等。通过引入万有引力理论、半监督学习理论、加权相似度理论等思想，提出了一系列改进的混合聚类模型，如基于万有引力理论的近邻传播聚类算法，基于加权相似度的自适应近邻传播聚类算法等。通过采用改进相似度、簇类中心等若干策略，提出了一系列具有更优聚类精度的聚类算法，有效提升了传统聚类模型的聚类性能，如基于改进距离函数的近邻传播聚类算法，基于截断距离和簇中心点自动选择策略的密度峰值聚类算法。