Data Noise Suppression-based Attribute Reduction

In recent years, as an effective tool for data preparation, attribute reduction can reduce the dimension. Meanwhile, the low dimensional data still can maintain a high accuracy to express the information of the original data. Therefore, attribute reduction has successfully attracted the attention of numerous scholars. However, varieties of noises exist in data, which leads to results of attribute reduction without a good generalization ability. To solve such a problem, a data noise suppression-based attribute reduction is proposed. Firstly, multiple data sets are obtained by multiple data acquisition. Secondly, multiple data sets are employed to generate a new data set. Thirdly, a new decision system is created over the new data set. Finally, the new decision system is used to derive reduct. To display the availability of the proposed strategy, the three classifiers (CART, KNN, SVM) and the measure of the approximation quality are employed in our experiments. Compared with the forward greedy searching strategy, the experiment results over 12 UCI data sets reveal that our proposed strategy can suppress additive noise in data. The generalization ability of reduct is significantly improved.