<p>In the real world, the presence of missing values brings a challenge to the classification, and missing value imputation is often inseparably involved in classification for incomplete data. In this paper, we propose a tracking-removed neural network with graph information for solving the classification problem of incomplete data. Specifically, we redesign the hidden layer neuron structure of the autoencoder to improve the network’s ability to mine associations among attributes. Graph information, which is used to analyze the similarity among samples, is introduced into the above tracking-removed neural network to further improve the network’s imputation performance for missing values. On the basis of appropriate imputation, the output layer neurons of the proposed network are reorganized to achieve the mapping of incomplete data to classification. Moreover, we present a learning algorithm that regards the missing values as variables and co-trains them with the network parameters for the designed model. The proposed strategy enables all the existing attribute information in incomplete datasets to participate in network training, which promotes the network to match the classification and regression structure of incomplete data, thereby improving the classification performance of the model for incomplete data. The experiments on 6 public datasets verify the effectiveness of the proposed method.</p> Graphical abstract <p></p>