Abstract
This paper introduces MI-GAN, a deep learning framework utilizing Generative Adversarial Networks (GANs) to address blockwise missing data in high-dimensional datasets. Unlike traditional methods like MICE that struggle with high dimensionality, or previous neural models limited to simple missingness assumptions, MI-GAN is theoretically proven to operate under the "missing at random" (MAR) mechanism. Empirical results demonstrate that MI-GAN is significantly faster and more statistically accurate than existing benchmarks, providing a scalable solution for complex data imputation.
Introduction
In modern data science, particularly within medicine and finance, researchers frequently merge datasets from disparate sources. This process often results in blockwise missingness, where specific variables are entirely absent for certain subsets of the data. For instance, one hospital may record a specific set of diagnostic tests while another records a different, partially overlapping set.
Traditional approaches to this problem are often inadequate. "Complete case analysis," which involves discarding rows with any missing values, introduces significant bias. Single imputation methods fail to account for the inherent uncertainty of missing data. While Multiple Imputation (MI) is considered the gold standard—generating multiple complete datasets to reflect uncertainty—existing tools like Multivariate Imputation by Chained Equations (MICE) are computationally prohibitive in high-dimensional settings where the number of features ($p$) exceeds the number of samples ($n$).
Methodology
MI-GAN addresses these challenges by leveraging the generative capabilities of GANs. The framework consists of two primary variants:
- MI-GAN₁ (Direct Imputation): This model employs an ensemble of $(K-1)$ GANs, corresponding to the number of missingness patterns ($K$). Each GAN includes a Generator ($G$) and a Discriminator ($D$). The generators are trained on complete cases to learn the conditional distribution of missing features given observed ones, utilizing the Wasserstein GAN (WGAN) framework with a gradient penalty for stability.
- MI-GAN₂ (Iterative Imputation): To handle scenarios where complete cases are scarce, MI-GAN₂ uses an iterative refinement process. It begins with an initial imputation and cycles through missingness patterns, using the entire dataset to progressively improve the generators.
The core innovation of MI-GAN lies in its ability to sample from the entire distribution of possible values rather than predicting a single mean. By varying the noise input to the generator, researchers can easily produce multiple imputations, making it a natural fit for the MI paradigm.
Experimental Results
The researchers evaluated MI-GAN against benchmarks including MICE, GAIN, and SoftImpute using both high-dimensional synthetic datasets ($p$ up to 1501) and real-world data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).
Scalability and Efficiency
A primary advantage of MI-GAN is its scalability. In synthetic tests with $p=1501$, MI-GAN₁ completed imputations in just 4.7 seconds. In contrast, state-of-the-art methods like Sinkhorn required nearly 148 seconds, while MICE failed entirely due to memory constraints.
Statistical Validity
Beyond error minimization (MSE), MI-GAN excels in statistical inference. It achieves a coverage rate (CR) near the nominal 95% level, ensuring that confidence intervals derived from the imputed data are reliable. On the ADNI dataset, MI-GAN produced regression coefficients ($\hat{\beta}_1$) that most closely aligned with the "Golden Standard" of complete data.
Conclusion and Future Work
MI-GAN provides a theoretically grounded and computationally efficient solution for blockwise missingness in high-dimensional data. While highly effective, the authors note that GANs require sufficient sample sizes for optimal convergence. Future research may explore the integration of Graph Neural Networks (GNNs) to better capture feature relationships, potentially enhancing performance in extremely high-dimensional or low-sample-size environments.
Summary updated