LG AI Research 550

The massive volumes of Tabular Data generated in manufacturing processes and inspection procedures are high-dimensional and heterogeneous in nature, making effective learning difficult with existing methods. Tree-based models such as the existing XGBoost had scalability issues when learning from large-scale data, and existing self-supervised learning (SSL) techniques were optimized for images or text, failing to reflect the structural differences of tabular data.

To overcome these limitations, LG AI Research proposed RaTab (Representation Space Augmentation for Tabular Data), a technology that performs data augmentation not in the Input Space but in the Representation Space. RaTab is designed to generate more meaningful representations by transforming the weight matrix of the encoder's final MLP layer.

This technology has the following differentiators:

• Overcoming limitations of existing methods: It prevents the distortion of relationships between key features and information loss caused by existing augmentation techniques such as Masking or Shuffling.
• Effective SSL execution: Augmentation in the representation space enables self-supervised learning optimized for tabular data.
• Real-world industrial application: It can effectively learn complex data such as sensor data and equipment logs from manufacturing sites, contributing to process optimization and quality control.