Meta-Learning Strategy and Its Performance
Evaluation
Yang Liying,Qin Zheng,Hu Guangwu, Zhang Xuanping
(School of Electronics and Information Engineering,Xi'an Jiaotong University,Xi'an
710049,China)
Abstract:Aiming at improving the classification performance, a meta-learning framework
covering many meta-learning algorithms was presented. Two modes of combining strategy,
parallel combining and serial combining,were defined and described in detail.The
classification results of basis classifiers were added to original feature vector to
obtain the meta-level data,which enhanced the search ability in hypothesis space and
decreased system bias.An experimental investigation was performed on UCI datasets and
encouraging results were obtained. Comparing with fusion methods of majority voting,max
rule, min rule etc,using the parallel and serial combined strategies proposed in this
paper the averaging error rate decreases by 39.12% and 40.56% respectively on the UCI
datasets applied.It is also shown that the increase of n in n-fold cross validation cannot
improve the classification performance, and the sequence of basis classifiers has no
significant affect on the error rate.
Keywords:pattern classification;multiple classifier system;meta-learning