Adaptive Graph Convolutional Network for Unsupervised Generalizable Tabular Representation Learning.

Area of Science:

• Deep Learning
• Machine Learning
• Data Representation

Background:

• Tabular data representation remains a challenge in deep learning.
• Existing autoencoder methods struggle to preserve discriminative information.
• Lack of effective neural architectures for tabular data structure exploration.

Purpose of the Study:

• Propose a novel adaptive graph convolutional network (AdaGCN) for unsupervised, generalizable tabular representation learning.
• Address limitations of current methods in capturing informative structures and preserving discriminative information.
• Develop a robust method for handling arbitrary tabular data.

Main Methods:

• Introduced an adaptive graph learning module, removing predefined rules for exploring local patterns.
• Employed an unsupervised approach minimizing distribution differences between original data and learned embeddings.
• Utilized a parametric property for efficient offline handling of unseen data.

Main Results:

• AdaGCN demonstrated significant and consistent performance improvements over existing representation learning and clustering methods.
• The adaptive graph learning module effectively explored local patterns on diverse tabular datasets.
• Unsupervised training preserved discriminative information, enhancing representation quality.

Conclusions:

• AdaGCN offers a powerful solution for unsupervised generalizable tabular representation learning.
• The adaptive and unsupervised nature of AdaGCN enhances its practicality and scope of applications.
• AdaGCN significantly advances the state-of-the-art in deep learning for tabular data.