TL;DR
This paper introduces Missingness Augmentation (MisA), a novel data augmentation technique for generative imputation models. MisA dynamically creates augmented incomplete samples during training, forcing the model to not only impute missing values but also to reconstruct observed ones, leading to significantly improved imputation accuracy across various datasets and models.
What Problem Does This Paper Address?
Missing data is a pervasive issue in real-world datasets, hindering the performance of machine learning models. While various imputation methods exist, they often struggle with complex data structures or limited training data. Data augmentation, a common technique to boost model performance by increasing data diversity, has been underutilized for missing data imputation. Existing augmentation methods often rely on inherent data structures (like pixel correlations in images) that are absent in tabular data, and they can struggle with high missing rates. This paper aims to bridge this gap by proposing a general, effective, and easy-to-integrate data augmentation strategy specifically designed for generative imputation models.
Methods
The core of the proposed method, Missingness Augmentation (MisA), is a dynamic data augmentation technique integrated into the training loop of generative imputation models. Instead of generating entirely new data, MisA leverages the generator's own output from a mini-batch. For each mini-batch, it:
- Generates an artificial mask: This mask mimics the original missingness pattern of the data, ensuring the augmented samples have a similar proportion of missing values.
- Creates augmented incomplete samples: It combines the generator's output with noise for the missing parts and uses the original observed values.
- Introduces a reconstruction loss: A new loss term, $L_{aug}$, is added. This loss penalizes discrepancies between the generator's imputed output for the augmented sample and the original observed values of that sample. This forces the generator to learn to reconstruct not just missing values but also to accurately represent the observed ones.
- Combines losses: The original imputation loss ($L_{orig}$) and the new reconstruction loss ($L_{aug}$) are combined into a single optimization objective with a regularization hyperparameter $\alpha$.
This approach is "plug-and-play," meaning it can be seamlessly integrated into existing generative imputation frameworks without altering their core architecture.
Datasets
The paper evaluates MisA on a variety of datasets:
- Image Datasets:
- MNIST
- CIFAR-10
- CelebA
- Tabular Datasets (from UCI Repository):
- Abalone
- Avila
- Ionosphere
- News
- Pendigits
- Sonar
- Wine
- WineQuality
No new datasets were introduced.
Key Insights and Ideas
- Imputed values should reflect raw observations: The paper's central hypothesis is that a good imputation model should not only fill in missing values but also ensure that its imputed values are consistent with the observed data. This is a crucial insight that guides the design of the reconstruction loss.
- Dynamic, generator-driven augmentation: MisA's novelty lies in using the generator's own output to create augmented samples. This makes the augmentation process adaptive to the model's current learning state.
- Generalizability across data types: The method is shown to be effective for both highly structured image data and heterogeneous tabular data, highlighting its broad applicability.
- Simplicity and efficiency: MisA requires only a simple reconstruction loss and minimal computational overhead, making it practical for real-world applications.
Key Results
- Significant performance improvements: MisA consistently improved the performance of several benchmark generative imputation models (DAE, DAEMA, GAIN, MCFlow, VAEAC, MisGAN) across various datasets.
- Tabular Data: On UCI datasets with a 50% missing rate, MisA achieved lower Root Mean Square Error (RMSE) compared to baseline models in the majority of cases. For example, GAIN+ showed a notable improvement over GAIN.
- Image Data: For image datasets (MNIST, CIFAR-10), MisA-augmented models achieved better PSNR and SSIM scores and lower FID scores, indicating higher quality imputed images. Figure 2 and 3 visually demonstrate that MisA-generated imputations are more cohesive and smoother.
- Robustness to missing rates: Table 4 shows that MisA's advantage often becomes more pronounced as the missing rate increases.
- Post-imputation accuracy: MisA also improved post-imputation prediction accuracy for classification tasks on both UCI and image datasets.
- Efficiency: The additional training time for MisA was generally small, around 10% for most models, demonstrating its efficiency.
Limitations and Future Work
- Missingness Distribution Mismatch: The paper notes that in MAR and MNAR scenarios, the missingness distributions of augmented samples can differ significantly from the actual missing mechanism. Constructing augmented samples that better conform to the specific missing mechanism could be a future direction.
- Hyperparameter Tuning: While MisA is generally robust to regularization coefficients ($\alpha$), finding the optimal value might still require some tuning.
- Future Work: The authors suggest that their approach can be a useful tool for addressing missing information in complex images and for deep generative imputation tasks in general, implying potential for further exploration in specific domains.
Summary updated