In the field of cancer classification, gene expression profile data has the characteristics of high dimensions, high redundancy, and unbalanced class distribution, which are the factors that affect the accuracy of classification. In order to improve the accuracy of cancer classification, this paper proposed a cancer classification method based on feature interaction and weight integration. At the feature selection level, this method used the gaining interaction of multiple features to select the features with the joint mutual information that was greater than the sum of the individual mutual information, and further used conditional mutual information to select low-redundancy features. At the classification model level, the re-learning ensemble model combined with weight integration feedback mechanism could comprehensively consider the different fitting ability of multiple models for different types of samples. This model constructed class weight that did not depend on the number of samples, and solved the problem of unbalanced class distribution. Comparative experiments of six kinds of cancer data show that the four indicators of accuracy, sensitivity, precision and F-measure are all stable above 99.39%, and the specificity is above 94.74%, which indicates that the method can further improve the accuracy and stability of cancer classification and the versatility of different cancers.

[2021-04-05] Printed Article /doc/2020.04.0106/v2