【摘要】：傳統(tǒng)特征選擇方法多是通過(guò)相關(guān)度量來(lái)去除冗余特征,并沒(méi)有考慮到高維相關(guān)矩陣中會(huì)存在大量的噪聲,嚴(yán)重地影響特征選擇結(jié)果。為解決此問(wèn)題,提出基于隨機(jī)矩陣?yán)碚?RMT)的特征選擇方法。首先,將相關(guān)矩陣中符合隨機(jī)矩陣預(yù)測(cè)的奇異值去除,從而得到去噪后的相關(guān)矩陣和選擇特征的數(shù)量;然后,對(duì)去噪后的相關(guān)矩陣進(jìn)行奇異值分解,通過(guò)分解矩陣獲得特征與類的相關(guān)性;最后,根據(jù)特征與類的相關(guān)性和特征之間冗余性完成特征選擇。此外,還提出一種特征選擇優(yōu)化方法,通過(guò)依次將每一個(gè)特征設(shè)為隨機(jī)變量,比較其奇異值向量與原始奇異值向量的差異來(lái)進(jìn)一步優(yōu)化結(jié)果。分類實(shí)驗(yàn)結(jié)果表明所提方法能夠有效提高分類準(zhǔn)確率,減小訓(xùn)練數(shù)據(jù)規(guī)模。
[Abstract]:The traditional feature selection method is to remove the redundant features through the correlation measure, and does not take into account that there is a large amount of noise in the high-dimensional correlation matrix, and the feature selection result is seriously affected. In order to solve this problem, a feature selection method based on random matrix theory (RMT) is proposed. firstly, removing the singular value of the correlation matrix according to the prediction of the random matrix, so as to obtain the number of the noise-removing correlation matrix and the selection characteristic; then, carrying out singular value decomposition on the noise-removing correlation matrix, and obtaining the correlation of the characteristic and the class through the decomposition matrix; and finally, Feature selection is done based on the redundancy between the features and the class's dependencies and features. In addition, a feature selection optimization method is propose, by which each feature is set to a random variable, and the difference between the singular value vector and the original singular value vector is compared to further optimize the result. The results of the classification experiment show that the proposed method can effectively improve the classification accuracy and reduce the training data scale.
【作者單位】： 遼寧大學(xué)信息學(xué)院;榮科科技股份有限公司智慧城市開(kāi)發(fā)部;
【基金】：國(guó)家自然科學(xué)基金資助項(xiàng)目(61472169,61472072,61528202,61501105) 國(guó)家973計(jì)劃前期研究專項(xiàng)(2014CB360509) 遼寧省教育廳科學(xué)研究一般項(xiàng)目(L2015204)~~
【分類號(hào)】：TP301.6
，

本文編號(hào)：2497362