Knowledge extraction from information stored in databases is always subject to the presence of missing values. Missing data is an unavoidable problem that affects many disciplines of researchers and data scientists. Inasmuch as machine learning algorithms cannot work with incomplete data in the data sets, data imputation is an essential task to obtain quality data. This research approach provides an overview of the data missingness mechanism and the process of generating synthetic missing data, the imputation of all types of variables, and the performance assessment of several imputation methods. Traditional algorithms, Machine Learning methods and various Autoencoder-based deep learning architectures have been studied. An exhaustive analysis and comparison of 21 heterogeneous data sets in various areas has been proposed. They have been exposed to a perturbation procedure with different missingness mechanisms and various missingness rates, covering the different possibilities that can occur in real life. The experimental results show that deep learning models outperform the other methods studied. Furthermore, the performance of data imputation methods does not depend on the missingness mechanism or the synthetic missingness generation method used nor on the percentage of missing values.