A Data Imputation Method Based on Deep Belief Network

Abstract

Accurately filling missing values is an important step to enhance the usability of Big Data. However current incomplete data imputation algorithms are of high time complexity and low accuracy. To address this problem, we propose a novel algorithm to impute incomplete data. Firstly, a deep belief network model with denoising is designed to remove the noise brought by incomplete data and extract high quality features. Then, we utilize softmax for data classification. Finally, according to the classification results, partial distance and sequence imputation strategies are proposed to measure the correlation between records and improve filling accuracy, respectively. Compared with different algorithms, the experimental results confirm the effectiveness and efficiency of the proposed method in data imputation.