Combining Distributed Classifies by Stacking

Many current mining tasks analyze data in environments with distributed computing nodes. Classification in such scenario needs to perform local mining task in each data site and then integrate local classifiers to a global model of the data. However, integration strategy can influence the performance and complexity of the final model. In this paper, based on the formalization of combining multiple classifiers by stacking in Distributed Data Mining, a new strategy to from meta-level training set is proposed, which can describe the vote made by each base-level classifiers. The experiment results show that our method achieve better performance for those datasets with highly skewed class distribution.