University of Coimbra, CISUC/LASI - Centre for Informatics and Systems of the University of Coimbra, Department of Informatics Engineering, 3030-290 Coimbra, Portugal
Auto-generated profile, based on publicly available information
Ricardo Cardoso Pereira
University of Coimbra, CISUC/LASI - Centre for Informatics and Systems of the University of Coimbra, Department of Informatics Engineering, 3030-290 Coimbra, Portugal
The performance evaluation of imputation algorithms often involves the generation of missing values. Missing values can be inserted in only one feature (univariate configuration) or in several features (multivariate configuration), at different percentages (missing rates) and according to distinct missing mechanisms, namely Missing Completely At Random, Missing At Random and Missing Not At Random. Since the missing data generation process defines the basis for the imputation experiments (configuration, missing rate, missing mechanism) it is essential that it is appropriately applied; otherwise, conclusions derived from ill-defined setups may be invalid. The goal of this work is to review the different approaches to synthetic missing data generation found in the literature and discuss their practical details, elaborating on their strengths and weaknesses. Our analysis revealed that creating Missing At Random and Missing Not At Random scenarios in datasets comprising qualitative features is the most challenging issue in related work and therefore should be the focus of future work in the field.
14 December 2020 in Journal of Artificial Intelligence Research
Missing data is a problem often found in real-world datasets and it can degrade the performance of most machine learning models. Several deep learning techniques have been used to address this issue, and one of them is the Autoencoder and its Denoising and Variational variants. These models are able to learn a representation of the data with missing values and generate plausible new ones to replace them. This study surveys the use of Autoencoders for the imputation of tabular data and considers 26 works published between 2014 and 2020. The analysis is mainly focused on discussing patterns and recommendations for the architecture, hyperparameters and training settings of the network, while providing a detailed discussion of the results obtained by Autoencoders when compared to other state-of-the-art methods, and of the data contexts where they have been applied. The conclusions include a set of recommendations for the technical settings of the network, and show that Denoising Autoencoders outperform their competitors, particularly the often used statistical methods.
Missing data can pose severe consequences in critical contexts, such as clinical research based on routinely collected healthcare data. This issue is usually handled with imputation strategies, but these tend to produce poor and biased results under the Missing Not At Random (MNAR) mechanism. A recent trend that has been showing promising results for MNAR is the use of generative models, particularly Variational Autoencoders. However, they have a limitation: the imputed values are the result of a single sample, which can be biased. To tackle it, an extension to the Variational Autoencoder that uses a partial multiple imputation procedure is introduced in this work. The proposed method was compared to 8 state-of-the-art imputation strategies, in an experimental setup with 34 datasets from the medical context, injected with the MNAR mechanism (10% to 80% rates). The results were evaluated through the Mean Absolute Error, with the new method being the overall best in 71% of the datasets, significantly outperforming the remaining ones, particularly for high missing rates. Finally, a case study of a classification task with heart failure data was also conducted, where this method induced improvements in 50% of the classifiers.