【摘要】：屬性效應(yīng)在現(xiàn)實(shí)生活中廣泛存在,如果不加以控制,將會(huì)嚴(yán)重影響回歸學(xué)習(xí)的性能.針對大規(guī)模數(shù)據(jù)屬性效應(yīng)控制的非線性回歸學(xué)習(xí)問題,提出了快速等均值核心向量回歸機(jī)(fast equal mean-core vector regression,FEM-CVR).首先基于間隔最大化目標(biāo)學(xué)習(xí)準(zhǔn)則,通過施加等均值約束條件,提出了等均值支持向量回歸機(jī)(equal mean-support vector regression,EM-SVR).在此基礎(chǔ)上,證明了EMSVR等價(jià)于一個(gè)中心約束最小包含球(center constrained-minimum enclosing ball,CC-MEB)問題,然后通過引入近似最小包含球理論,得到原始輸入數(shù)據(jù)集的壓縮集即核心集(core set),進(jìn)一步提出了針對大規(guī)模數(shù)據(jù)屬性效應(yīng)控制的最小包含球快速非線性回歸學(xué)習(xí)方法 FEM-CVR,并從理論上對相關(guān)性質(zhì)進(jìn)行了深入分析.實(shí)驗(yàn)表明:該方法能夠快速處理針對大規(guī)模數(shù)據(jù)屬性效應(yīng)控制的非線性回歸學(xué)習(xí)問題,具有較好的泛化能力,并且其時(shí)間復(fù)雜度上限與數(shù)據(jù)集大小無關(guān),僅與最小包含球近似參數(shù)ε-有關(guān).
[Abstract]:Attribute effect exists widely in real life, if it is not controlled, it will seriously affect the performance of regression learning. A fast equal-mean kernel vector regression machine (fast equal mean-core vector regression,FEM-CVR) is proposed for nonlinear regression learning of attribute effect control in large-scale data. Based on the goal learning criterion of maximizing the interval, the equal-mean support vector regression machine (equal mean-support vector regression,EM-SVR) is proposed by applying the equal mean constraint condition. On this basis, it is proved that EMSVR is equivalent to a central constrained minimum inclusion sphere (center constrained-minimum enclosing ball,CC-MEB) problem. Then, by introducing the theory of approximate minimum inclusion sphere, the compressed set of the original input data set, that is, the core set (core set), is obtained. Furthermore, a fast nonlinear regression learning method, FEM-CVR, is proposed for attribute effect control of large scale data, and the related properties are analyzed theoretically. Experiments show that this method can deal with the nonlinear regression learning problem for large-scale data attribute effect control quickly, and has good generalization ability, and the upper limit of time complexity is independent of the size of data set. It is only related to the minimal inclusion sphere approximation parameter 蔚 -.
【作者單位】： 江南大學(xué)數(shù)字媒體學(xué)院;湖北交通職業(yè)技術(shù)學(xué)院交通信息學(xué)院;
【基金】：國家自然科學(xué)基金項(xiàng)目(61300151,61572236) 江蘇省杰出青年基金項(xiàng)目(BK20140001) 江蘇省自然科學(xué)基金項(xiàng)目(BK20130155,BK20151299)~~
【分類號(hào)】：TP181

【相似文獻(xiàn)】

相關(guān)期刊論文 前2條

1 羅佳;薛青;唐志武;;作戰(zhàn)仿真數(shù)據(jù)屬性離散化研究[J];計(jì)算機(jī)仿真;2013年09期

2 胡峰;張杰;劉靜;肖大偉;;一種基于Rough集的海量數(shù)據(jù)屬性約簡方法[J];重慶郵電大學(xué)學(xué)報(bào)(自然科學(xué)版);2009年04期

，

本文編號(hào)：2372950