Graph Neural Network contextual embedding for Deep Learning on tabular data.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Industries utilize big data in tabular format, comprising heterogeneous features.
• Deep Learning (DL) excels in areas like natural language processing but faces challenges with tabular data.
• Classical Machine Learning (ML) models, particularly tree-based ensembles, often outperform DL on tabular datasets.

Purpose of the Study:

• To present a novel Deep Learning (DL) model for tabular data analysis.
• To leverage Graph Neural Networks (GNNs), specifically Interaction Networks (INs), for contextual embedding and feature interaction modeling.
• To demonstrate the model's effectiveness against existing DL benchmarks and traditional ML models.

Main Methods:

• Development of a novel DL model based on Graph Neural Network (GNN) architecture.
• Utilization of Interaction Network (IN) for contextual embedding of tabular features.
• Evaluation on seven public datasets against a DL benchmark and boosted-tree solutions.

Main Results:

• The proposed GNN-based model outperforms a recent DL benchmark on tabular data.
• The model achieves competitive performance when compared to established boosted-tree ML solutions.
• Demonstrates improved modeling of interactions among heterogeneous tabular features.

Conclusions:

• Graph Neural Networks (GNNs), specifically Interaction Networks (INs), offer a promising DL approach for tabular data.
• The novel model provides a viable alternative to traditional ML methods for tabular data analysis.
• This research advances the application of DL techniques to complex, real-world tabular datasets.