如何有效进行数据归约是大规模数据挖掘的一个关键问题。嵌入式数据归约是指在学习数据挖掘任务的同时进行数据归约，是当前数据归约的研究热点。支持向量机是嵌入式数据归约的典型代表，然而传统的支持向量机只能进行样本归约，不能进行特征归约，并且对大规模数据的归约性能也亟待提高。本项目拟进一步改进支持向量机及有关嵌入式数据归约方法，施行大规模嵌入式数据归约。具体内容包括：研究对样本和特征具有更加稀疏性质的损失项和正则项，进而提出即可进行样本规约、又可进行特征归约的大规模支持向量机；对最小二乘支持向量机和判别分析，构造稀疏的特征选择和特征提取模型与算法；研究可以同时进行样本和特征归约的稀疏模型，重点关注其稀疏正则化和稀疏损失函数的可解释性和大规模问题的快速求解算法；最后，将以上研究应用于大规模图像信息检索问题。本项目的研究对大规模数据挖掘中嵌入式数据归约的理论和实践都具有创新意义。

How to effectively perform data reduction is a key issue for large-scale data mining. Embedded data reduction refers to data reduction while learning data mining tasks. It is a research hotspot for current data reduction. Support vector machine (SVM) is a typical representative of embedded data reduction. However, traditional support vector machines can only perform sample reduction, and cannot perform feature reduction. Furthermore, it is particularly important data to improve the performance of the data reduction for the large-scale. This project intends to further improve support vector machines and related embedded data reduction methods for large-scale embedded data reduction. The specific contents include: Study loss functions and regular terms with more sparse nature of both samples and features, and then propose the large-scale support vector machines that can perform sample reduction and feature reduction simultaneously; For least squares SVM and linear discriminant analysis, study the sparse models and algorithms for feature selection and feature extraction; study the sparse models that can simultaneously perform sample and feature reduction, focusing on the interpretability of sparse regularization and sparse loss functions, and fast algorithms of large-scale problems Finally, the above research is applied to the large-scale image information retrieval problem. The study of this project is of innovative significance to the theory and practice of embedded data reduction in large-scale data mining.