Missing data imputation with adversarially-trained graph convolutional networks.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Missing data is a pervasive issue across scientific disciplines, necessitating robust imputation methods.
• Existing missing data imputation (MDI) techniques often rely on global statistics or instance-independent models, limiting their effectiveness.
• Graph neural networks (GNNs) offer a powerful paradigm for learning complex data relationships.

Purpose of the Study:

• To propose a generalized framework for missing data imputation using graph neural networks.
• To leverage graph denoising autoencoders for reconstructing datasets with missing values.
• To enhance imputation accuracy and efficiency through advanced architectural and loss function choices.

Main Methods:

• Formulated MDI as a graph denoising autoencoder problem, encoding pattern similarity as graph edges.
• Employed a GNN encoder-decoder architecture to learn intermediate representations and reconstruct the full dataset.
• Utilized a combination of losses, including Wasserstein adversarial loss with gradient penalty, and explored residual connections and global statistics.

Main Results:

• The proposed GNN-based MDI method achieved performance on par with or superior to existing methods on datasets with artificial noise.
• Evaluations on datasets with naturally occurring missing values demonstrated the method's robustness.
• The imputation approach yielded comparable or improved results with downstream classifiers.

Conclusions:

• The GNN framework provides a flexible and effective approach to missing data imputation.
• This method demonstrates strong performance and robustness compared to traditional imputation techniques.
• The study highlights the potential of GNNs for addressing fundamental data challenges in various scientific fields.