PC-GAIN: Pseudo-label conditional generative adversarial imputation networks for incomplete data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Real-world datasets frequently contain missing values, necessitating effective imputation techniques.
• Existing deep generative models like GAIN (Generative Adversarial Imputation Nets) show promise but overlook sample relationships.
• GAIN's reconstruction loss alone may not fully capture data structure, limiting imputation performance.

Purpose of the Study:

• To propose PC-GAIN, a novel unsupervised method for missing data imputation.
• To enhance the imputation power of deep generative models by incorporating potential category information.
• To improve the quality of imputation results compared to existing state-of-the-art methods.

Main Methods:

• PC-GAIN utilizes a pre-training strategy to identify potential category information from low-missing-rate data subsets.
• An auxiliary classifier is trained using synthetic pseudo-labels derived from the pre-training phase.
• The trained classifier is integrated into the generative adversarial framework to guide the generator towards higher-quality imputations.

Main Results:

• The proposed PC-GAIN method demonstrates significant improvements in imputation quality compared to the original GAIN model.
• Experimental results on benchmark datasets confirm PC-GAIN's superiority over other baseline imputation approaches.
• The integration of category information effectively enhances the performance of deep generative imputation.

Conclusions:

• PC-GAIN offers a powerful enhancement for deep generative missing data imputation by utilizing potential category information.
• The method provides a novel approach to unsupervised imputation, improving accuracy and data representation.
• The findings suggest that incorporating sample relationships through category information is crucial for robust missing data handling.