基于矩阵分解优化的排序学习特征构造方法

发布时间：2018-04-07 17:08

本文选题：特征生成　切入点：排序学习　出处：《计算机科学》2017年12期

【摘要】：在排序学习中引入特征选择可以提高学习的效率和准确率。出于对选择速度的考虑,当前的研究主要从特征选择的角度出发,根据特征对排序的作用和特征之间的相似性选择对排序区分度最大的特征集合。由于特征大都是人工归纳的,因此特征和特征之间难免存在重叠和冗余。为了减少特征之间的冗余,从特征生成的角度出发,对现有特征进行矩阵分解,从而生成新的特征集。考虑到使用奇异值分解(Singular Value Decomposition SVD)等方法进行矩阵分解时不能综合考虑排序结果对特征的影响,基于特征矩阵对排序的效果、特征矩阵与原矩阵之间的差距来构造优化算法,提出了一种基于矩阵分解的排序学习优化方法,并根据该优化方法设计了排序学习特征选择算法MFRank。实验中使用映射随机梯度下降法近似求得优化问题的最优值,在公开测试集MQ2008上的结果显示,所提MFRank方法获得了与当前最优的特征选择方法即RankBoost和RankSVM-Struct等排序算法相当的结果。
[Abstract]:Feature selection in ranking learning can improve the efficiency and accuracy of learning.Considering the selection speed, the current research mainly focuses on the feature selection. According to the function of feature to sort and the similarity between features, the feature set with the largest degree of sorting is selected.Since most features are induced manually, there is inevitably overlap and redundancy between features and features.In order to reduce the redundancy between features, the matrix decomposition of existing features is carried out from the point of view of feature generation, and a new feature set is generated.Considering that the singular Value Decomposition (SVD) method can not synthetically consider the influence of the sorting results on the feature, the optimization algorithm is constructed based on the effect of the feature matrix on the sorting and the gap between the feature matrix and the original matrix.A sorting learning optimization method based on matrix decomposition is proposed, and MFRank-based sorting learning feature selection algorithm is designed.In the experiment, the mapping stochastic gradient descent method is used to approximate the optimal value of the optimization problem. The results on the open test set MQ2008 show that the proposed MFRank method is comparable to the current optimal feature selection methods, such as RankBoost and RankSVM-Struct.
【作者单位】：山东财经大学管理科学与工程学院;山东财经大学计算机科学与技术学院;齐鲁工业大学金融学院;曼彻斯特大学曼彻斯特商学院;
【基金】：国家自然科学基金项目:基于机器学习融合精确性和多样性的电子商务协同过滤推荐方法研究(71402083) 山东省高等学校科技计划项目:基于语义角色主题模型的细粒度情感分析研究(J15LN56)资助
【分类号】：TP181

【相似文献】