CACTI: Leveraging Copy Masking and Contextual Information to Improve Tabular Data Imputation

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Tabular data imputation is crucial for data analysis and machine learning.
• Existing methods often struggle with complex missingness patterns and lack of feature context.
• Effective imputation enhances downstream task performance and data integrity.

Purpose of the Study:

• To develop a novel masked autoencoding approach for tabular data imputation.
• To leverage both missingness patterns and contextual feature information for improved imputation.
• To outperform existing state-of-the-art imputation methods.

Main Methods:

• CACTI (Contextual Autoencoding for Tabular Data Imputation) employs masked autoencoding.
• A novel median truncated copy masking strategy is utilized during training.
• Semantic relationships from column names and descriptions are incorporated to model feature dependence.

Main Results:

• CACTI demonstrates superior performance across diverse datasets and missingness conditions.
• Achieved an average R-squared gain of 7.8% over the next best method.
• Outperformed methods by 13.4%, 6.1%, and 5.3% under missing not at random, at random, and completely at random scenarios, respectively.

Conclusions:

• Leveraging dataset-specific contextual information and missingness patterns enhances imputation performance.
• CACTI offers a robust and effective solution for tabular data imputation.
• The proposed method provides a valuable tool for data scientists and machine learning practitioners.