Incomplete Data Classification Based on the Tracking-Removed Autoencoder

The existence of missing values increases the difficulty of classification in the real world. In this paper, we propose a classification method based on the tracking-removed autoencoder (TRAE) for incomplete data. Specifically, we use the tracking-removed autoencoder as an imputation model to estimate missing values by constructing mutual fitting relationships between attributes on a network structure. On this basis, the multi-layer perceptron (MLP) is used as the classification model to classify the complete data after imputation. In the training phase, we designed a new training scheme for the imputation model, which treats missing values as variables and dynamically optimizes them together with network parameters, making it easier for missing values to match existing attribute values in incomplete datasets. Then, we use the complete training set after imputation to train the parameters in the MLP network model. In the testing phase, the network parameters of the imputation model are fixed and the missing values in the test set are considered as variables, which are optimally updated to obtain the complete test set. The classification model predicts its class label based on the optimized complete test set. The training scheme make the missing values in incomplete data approach the actual value with higher precision, and weaken the impact of missing data on classification. Experimental results on several incomplete datasets demonstrate that the proposed method can effectively improve the imputation performance, thereby improving the classification performance of incomplete data.