Deep Tabular Representation Corrector.

Area of Science:

• Machine Learning
• Data Science
• Deep Learning

Background:

• Tabular data is crucial in fields like healthcare and finance.
• Deep learning models (e.g., Transformer, ResNet) are increasingly used for tabular data.
• Existing methods like in-learning and pre-learning have limitations in efficiency and complexity.

Purpose of the Study:

• To introduce a novel deep Tabular Representation Corrector (TRC) to enhance representations of existing deep tabular models.
• To address representation shift and redundancy issues hindering prediction accuracy.
• To achieve model-agnostic enhancement without altering original model parameters.

Main Methods:

• Developed TRC, a model-agnostic approach to improve deep tabular model representations.
• Proposed Tabular Representation Re-estimation to mitigate inherent representation shifts.
• Introduced Tabular Space Mapping to create compact, informative embedding vectors, minimizing redundancy.

Main Results:

• TRC enhances representations by addressing shift and redundancy.
• The method is model-agnostic, improving various deep tabular models.
• Extensive experiments on benchmarks show consistent superiority of TRC-enhanced models.

Conclusions:

• TRC offers an efficient way to boost deep tabular model performance.
• The proposed re-estimation and mapping tasks effectively improve data representations.
• TRC represents a significant advancement in deep learning for tabular data analysis.