Ensemble Of Oselm For Large Data Sets Classification

With the rapid development of computer technology and the Internet, data is increasing at an unprecedented speed. Large number of data has been accumulated in many practical application fields. It is a hot research topic to improve the existed algorithms to adapt classification of large data sets.This paper investigates the problem of large data sets classification with OSELM(Online Sequential Extreme Learning Machine).OSELM algorithm is a sequential learning algorithm. When compared with other state-of-the-art sequential algorithms, OSELM has faster learning speed with very good generalization ability. But for large data sets, OSELM will not halt when there are training samples not be learned, this phenomenon results in long learning time. Moreover, OSELM also has instability in different trials of simulations. In order to deal with this problem, two algorithms for classifying the large data sets are proposed in this paper. Two algorithms are all composed of three steps.(1)Firstly, the component OSELM classifiers are sequentially trained on subsets of a large data set, in the process of training component classifiers, the instances previously used will be excluded from training the following component classifiers.(2) Secondly, the trained component classifiers are combined with fuzzy integral.(3) Finally, the aggregation learning system is used for classifying the unseen samples.The difference between the two algorithms is that they use different ensemble strategies to integrate the component classifiers. The first one uses a static integration strategy; the second algorithm uses a dynamic integration strategy. We compared our methods with two other state-of-the-art large data sets classification methods, which are DTSVM(Decision Tree Support Vector Machine) and CVM(Core Vector Machine). The experimental results show that the proposed method outperforms DTSVM and CVM, moreover the proposed approach can overcome instability of OSELM in different trials of simulations.