TabMixer: advancing tabular data analysis with an enhanced MLP-mixer approach.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Tabular data is crucial in various domains like healthcare, engineering, and finance.
• Existing tabular data learning methods face challenges with missing values, class imbalance, and transfer/incremental learning.

Purpose of the Study:

• Introduce TabMixer, an enhanced multilayer perceptron (MLP) mixer model.
• Address key challenges in tabular data analysis, including missing values and class imbalance.
• Enable versatile learning scenarios: supervised, transfer, and feature incremental learning.

Main Methods:

• Developed TabMixer, integrating a self-attention mechanism into the MLP mixer architecture.
• Evaluated TabMixer on eight public datasets across different learning paradigms.
• Assessed computational efficiency, scalability, and resilience to data imperfections.

Main Results:

• TabMixer demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements in ANOVA AUC: 4% in supervised learning, 8% in transfer learning, and 4% in feature incremental learning.
• Showcased high computational efficiency, scalability, and robustness to missing values and class imbalance.

Conclusions:

• TabMixer is a highly effective and versatile model for tabular data analysis.
• The model offers significant advantages in supervised, transfer, and feature incremental learning scenarios.
• TabMixer presents a promising tool for diverse real-world applications requiring robust tabular data processing.