元學(xué)習(xí)算法選擇機(jī)制及關(guān)聯(lián)對(duì)性能的影響

提出一種元學(xué)習(xí)定義，從偏差/方差分解角度對(duì)元學(xué)習(xí)中學(xué)習(xí)算法的選取機(jī)制進(jìn)行研究，得出了元級(jí)選用錯(cuò)誤率低且偏差小的學(xué)習(xí)算法、基級(jí)學(xué)習(xí)算法按照錯(cuò)誤率及方差從低到高排列的結(jié)論。鑒于標(biāo)準(zhǔn)數(shù)據(jù)集不能充分評(píng)估關(guān)聯(lián)對(duì)元學(xué)習(xí)性能的影響，設(shè)計(jì)了一種模擬算法以產(chǎn)生模擬數(shù)據(jù)集。在UCI標(biāo)準(zhǔn)數(shù)據(jù)集和模擬數(shù)據(jù)集上的實(shí)驗(yàn)表明，同常用的多數(shù)投票等組合方法相比，基于算法選擇機(jī)制的元學(xué)習(xí)表現(xiàn)出優(yōu)良的性能，且分類器之間的負(fù)關(guān)聯(lián)有助于性能的改進(jìn)。

In this paper, a general definition of meta-learning is proposed. The selection of learning algorithms in meta-learning is investigated from the point of bias/variance decomposition as well as the effect of correlation on its accuracy. In order to obtain classifiers with variable correlation, artificial datasets are generated based on the simulating algorithm presented in the paper. Experiments are performed on UCI datasets and simulated datasets and show that meta-learning outperforms several combining methods averagely; and that negative correlation measured by Q statistic benefits meta-learning approach.