Dual Joint Covariance Alignment Method for Incomplete Data Classification

For incomplete data classification, the missing values are usually imputed by different estimation methods to make the data complete. In practice, the estimated values are not real attribute values, so the distributions of imputed training and imputed test data are different. Here, the distribution discrepancy should be effectively reduced for further classification performance improvement. To this end, we present a dual-joint covariance alignment (DJCA) method for incomplete data classification. In DJCA, the imputation process is first employed to estimate missing values. Then, the global and local covariance differences between the imputed training and imputed test data are both reduced to learn two feature transformation matrices. One makes the distribution of imputed training data with transformation close to that of imputed test data, and the other one makes the distributions between transformed imputed test data and imputed training data similar. For a query pattern, it will obtain two pieces of soft classification results yielded by the auxiliary of the two matrices. The weighted arithmetic average (WAA) rule is then employed to integrate the soft classification results. Their weights are automatically learned by minimizing the mean squared error between WAA fusion results and ground truth on the labeled training data. DJCA was tested on several datasets, and the experimental results demonstrate that DJCA can significantly improve the classification accuracy compared to related methods (e.g., if the number of missing attributes is 3, and the base classifier is support vector machine, DJCA can achieve 4.63% higher classification accuracy on the Bupa dataset and 8.23% on the vertebral dataset compared to the classical mean imputation (MI) method).