超图低秩属性选择多输出回归算法

发布时间：2019-02-23 13:17

【摘要】：随着科技的发展,不少学者已经提出了许多方法来应对越来越具有挑战性的多输出回归任务。在高维数据的回归任务中,多输出回归算法利用数据的多个特征值来预测多个输出值,能获得较好的效果。实际应用中产生的数据往往具有很多的特征,比如生物医学的基因数据,计算机视觉中的图像数据,web网站中产生的日志或用户浏览记录的文本数据等。为了提高对这些高维数据的利用效率,并在回归分析中能取得比较好的分类效果,因此需要对这些高维数据进行特征选择,降低这些高维数据的维度。许多研究者提出了低秩回归模型,并在此基础上进行特征选择和属性约减。由于只是简单的进行处理,没有对样本进行选择(去除噪音数据和离群点数据对模型的干扰)。因此,为了能够快速且稳定地处理高维数据多输出回归问题,本文提出了一种超图低秩属性选择多输出回归算法。具体地,首先,本文在基本线性回归基础上应用l2,p范数进行样本选择,去除噪音和离群点样本,使得学习得到的模型具有更准确的预测能力。接着,通过回归系数矩阵来进行特征选择,从而得到新的特征样本数据集。然后,使用10-折交叉验证进行模型训练,从而学习最终回归模型。模型中学习得到的回归系数矩阵可以选择出重要的特征,而低秩回归可以输出各变量之间紧密的关联结构。最后,在模型中嵌入超图,通过超图来保留各数据之间的局部相关性结构,使得模型更具有稳定性。实验结果表明,该算法相比其它对比算法,能更有效地选取属性,应用于多输出回归任务中能取得很好的分类效果。
[Abstract]:With the development of science and technology, many scholars have proposed many methods to deal with the more and more challenging multi-output regression task. In the task of regression of high-dimensional data, multi-output regression algorithm uses multiple eigenvalues of data to predict multiple output values, which can obtain better results. The data produced in practical applications often have many characteristics, such as biomedical genetic data, image data in computer vision, logs produced in web websites or text data recorded by users. In order to improve the utilization efficiency of these high-dimensional data and obtain better classification results in regression analysis, it is necessary to select the features of these high-dimensional data and reduce the dimensions of these high-dimensional data. Many researchers have proposed a low rank regression model based on which feature selection and attribute reduction are carried out. Because of the simple processing, the sample is not selected (noise data and outlier data interference to the model). Therefore, in order to deal with the multi-output regression problem of high-dimensional data quickly and stably, this paper proposes a multi-output regression algorithm for hypergraph with low-rank attribute selection. Specifically, first of all, on the basis of basic linear regression, this paper uses L2, p norm to select samples to remove noise and outliers, so that the model obtained from learning has more accurate prediction ability. Then, the new feature sample data set is obtained by regression coefficient matrix. Then, the 10-fold cross-validation is used to train the model to learn the final regression model. The regression coefficient matrix obtained from the model can select the important characteristics, while the low rank regression can output the tight correlation structure among the variables. Finally, the hypergraph is embedded in the model, and the local correlation structure between the data is preserved by the hypergraph, which makes the model more stable. The experimental results show that the proposed algorithm can select attributes more effectively than other comparison algorithms, and can achieve good classification effect in multi-output regression task.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：O212.1

【相似文献】