基于数据特性分析的机器学习建模与算法研究

Machine leaning is the core supporting technology for big data analysis and processing, so further researches on machine learning modeling and algorithms for big data have important theoretical significance and application value. Aiming at some challenges in machine learning such as large scale and poor quality of data, the model constructing, learning and measurements based on data characteristic analysis will be studied deeply and systematically. The main contents of the project include: (1) Instruction of penalty factor in regularization machine learning models, and analysis of error bounds theory and properties. (2) Uniform framework of kernel selection based on orthogonal polynomial. (3) Estimation of generalization performance of algorithms by estimating cross-validation variance. (4) Analysis of relationships and pseudo relevance embedding in data. (5) Developing the high efficient machine learning algorithms for some special applications. The research results will not only enrich the related theories and algorithms of big data machine learning, but provide the important technical supports for practical applications of machine learning in big data as well.

机器学习是大数据分析与处理的核心支撑技术，深入研究大数据机器学习建模与算法具有重要的理论意义与应用价值。本项目针对数据的大规模、低质量等给机器学习带来的挑战，基于数据的特性分析，就模型构建、模型学习、模型评价等机器学习核心科学问题开展系统深入的研究。主要内容包括：（1）正则化机器学习模型中正则项的构造及其性质；（2）基于正交多项式的核函数构造统一框架；（3）基于交叉验证方差估计的学习算法泛化误差估计；（4）数据的相关性分析；（5）高效机器学习算法设计及应用。项目研究成果将丰富大数据机器学习的相关理论与算法，为机器学习的大数据实际应用提供重要的技术支撑。

本项目围绕大数据的分类、聚类、预测等数据分析和挖掘任务，针对复杂数据的大规模、低质量给传统机器学习方法带来的挑战，就机器学习的模型构建、模型学习、模型评价等机器学习核心科学问题开展了系统深入的研究，通过对数据进行特性分析，构建了有效的学习模型和高效的学习算法，并应用于服务计算、社会网络等领域。项目研究成果丰富了大数据机器学习的相关理论与算法，为机器学习的大数据实际应用提供了重要的技术支撑。项目组发表多篇高水平学术成果，培养多名博士、硕士研究生。项目研究成果已达到研究目标。