Incomplete data modeling based on alternate update of clustering and autoencoder for missing value imputation

Missing values exist widely in real-world datasets, which restrict the performance of data mining. In this paper, we propose a joint optimization framework to mine attribute associations and category structures in incomplete datasets, aiming to impute missing values with a full understanding of the data structure. Considering the differences in attribute correlations among different sample categories, we partition incomplete data into fuzzy subsets by fuzzy clustering. Within each subset, a tracking-removed autoencoder is constructed as a submodel to fit the regression relationships among attributes. Due to the mutual influence between fuzzy clustering and regression modeling, we further propose a missing value variable-based training scheme to iteratively optimize these two processes. Our proposed framework offers the advantage of decomposing the complex imputation task into simpler sub-tasks by fuzzy clustering where the attribute associations are more explicit. Moreover, the proposed training scheme activates the complementary nature of clustering and regression processes to reduce imputation errors. The experimental results on artificial and real datasets illustrate the effectiveness of our proposed framework.