Exploiting Field Dependencies for Learning on Categorical Data.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Traditional methods for categorical data learning often overlook inter-field dependencies.
• Existing approaches rely heavily on classification/regression loss for data point embedding, limiting performance.

Purpose of the Study:

• To develop a novel method for learning on categorical data that effectively exploits dependencies between fields.
• To improve the modeling of field dependencies through both global and local approaches.

Main Methods:

• Learns a global field dependency matrix to capture inter-field relationships.
• Refines the global matrix at an instance-wise level using local dependency modeling.
• Employs a meta-learning paradigm where dependency matrices are updated without labels in the inner loop and with labels in the outer loop.

Main Results:

• The proposed method significantly outperforms several state-of-the-art approaches on six benchmark datasets.
• Ablation studies confirm the effectiveness and provide insights into the method's components.

Conclusions:

• The novel meta-learning-based approach enhances categorical data learning by effectively modeling field dependencies.
• This method offers a simple yet powerful alternative to existing techniques, demonstrating superior performance.