BIWE: BOOSTING-BASED ITERATIVE WEIGHTED ENSEMBLE CLASSIFICATION

Classification is a function that plays a major role in the development of data mining, across a widespread variety of application fields: using classification accuracy and the scale of spanning tree are usually prime requirement. Ensemble classification is the most used method to deal with dynamic data streams problems in recent years. Boosting is renowned method to acquire more diversity and robust classifier from poor performance base classifiers, thus has been widely studied. However, common ensemble algorithms can only replace one base classifier at a time and cannot quickly restore the overall performance. Meanwhile, it is mandatory for the base classifier to use the same distribution weight method in different data sets, and it is too coercive. In this paper, we propose a new algorithm called Boosting-based Iterative Ensemble classification (BIE), which can calculate the number of base classifiers to be replaced in each iteration according to the classification accuracy of the latest incoming data streams. In addition, basis on BIE algorithm, we proposed Boosting-based Iterative Weighted Ensemble classification (BIWE) algorithm, which can calculate the optimal value corresponding to the weight of base classifier for data streams with different parameter characteristics. In order to better observe the performance of algorithms, we compare with 9 algorithms on 9 dynamic data streams. Experimental results show that BIE and BIWE algorithms not only have ideal classification accuracy, but also can greatly reduce the scale of spanning trees, in the aspect of the depth, the number of nodes and leaves of trees.