Open AccessWeighted ensemble based on 0‐1 matrix decomposition
Author(s) -
Mao Shasha,
Xiong Lin,
Jiao L.C.,
Zhang Shuang,
Chen Bo
Publication year - 2013
Publication title -
electronics letters
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.375
H-Index - 146
ISSN - 1350-911X
DOI - 10.1049/el.2012.3528
Subject(s) - singular value decomposition , mathematics , singular value , matrix (chemical analysis) , weight , matrix decomposition , simple (philosophy) , decomposition , value (mathematics) , ensemble learning , algorithm , pattern recognition (psychology) , artificial intelligence , statistics , computer science , eigenvalues and eigenvectors , pure mathematics , physics , ecology , philosophy , materials science , epistemology , quantum mechanics , lie algebra , composite material , biology
A simple effective ensemble method is proposed in which individual classifiers are combined with the weight coefficients obtained by decomposition for the 0‐1 matrix. The 0‐1 matrix is introduced to denote individual classifiers of the ensemble and is constructed based on the prediction labels of individuals and the true labels. The weight coefficients of individuals are obtained by singular value decomposition for the 0‐1 matrix based on linear mapping. In particular, the square of elements of the right singular vector corresponding to the maximum singular value are as the weight coefficients of individuals, and it is proved theoretically that it minimises the upper bound of ensemble error. Experimental results illustrate that the proposed method improves the performance of classification compared against standard ensemble strategies.