Multiple-Side Multiple-Learner for Incomplete Data Classification

Selective classifier can improve classification accuracy and algorithm efficiency by removing the irrelevant attributes of data. However, most of them deal with complete data. Actual datasets are often incomplete due to various reasons. Incomplete dataset also have some irrelevant attributes which have a negative effect on the algorithm performance. By analyzing main classification methods of incomplete data, this paper proposes a Multiple-side Multiple-learner algorithm for incomplete data (MSML). MSML first obtains a feature subset of the original incomplete dataset based on the chi-square statistic. And then, according to the missing attribute values of the selected feature subset, MSML obtains a group of data subsets. Each data subset was used to train a sub classifier based on bagging algorithm. Finally, the results of different sub classifiers were combined by weighted majority voting. Experimental results on UCI incomplete datasets show that MSML can effectively reduce the number of attributes, and thus improve the algorithm execution efficiency. At the same time, it can improve the classification accuracy and algorithm stability too.