Towards Better Modeling With Missing Data: A Contrastive Learning-Based Visual Analytics Perspective.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Missing data presents a significant challenge for machine learning (ML) model performance.
• Current methods for handling missing data, such as feature imputation and label prediction, have limitations including distribution assumptions and potential bias.
• These traditional approaches require imputation, which can be complex and introduce inaccuracies.

Purpose of the Study:

• To propose a novel Contrastive Learning (CL) framework for modeling data with missing values.
• To address the shortcomings of traditional imputation methods in ML.
• To enhance both the predictive accuracy and interpretability of ML models dealing with incomplete datasets.

Main Methods:

• Developed a Contrastive Learning (CL) framework to model observed data with missing values.
• The CL model learns similarities between incomplete and complete samples, and dissimilarities among other samples.
• Introduced CIVis, a visual analytics system for enhanced interpretability, enabling users to interactively identify positive and negative pairs.

Main Results:

• The proposed CL framework effectively handles missing data without requiring imputation.
• Achieved high predictive accuracy in both regression and classification tasks.
• Demonstrated enhanced model interpretability through the CIVis visual analytics system.

Conclusions:

• The Contrastive Learning (CL) framework offers a practical and effective solution for ML modeling with missing data.
• The approach overcomes limitations of traditional imputation methods, providing high accuracy and interpretability.
• CIVis enhances user understanding and control over the CL process for missing data imputation.