CDE++: Learning Categorical Data Embedding by Enhancing Heterogeneous Feature Value Coupling Relationships.

Area of Science:

• Machine Learning
• Data Representation

Background:

• Categorical data is common in machine learning.
• Effective representation is crucial for performance.
• Real-world data has complex, heterogeneous feature couplings.

Purpose of the Study:

• To propose an enhanced categorical data embedding method, CDE++.
• To capture heterogeneous feature value coupling relationships.
• To improve machine learning performance on categorical data.

Main Methods:

• CDE++ builds on CDE (Categorical Data Embedding).
• Utilizes information theory: mutual information and margin entropy for feature couplings.
• Employs a hybrid clustering strategy for feature value clusters.
• Incorporates Autoencoder for non-linear couplings.

Main Results:

• CDE++ generates low-dimensional numerical vector embeddings.
• Embeddings are directly applicable to clustering and classification.
• Achieved superior performance compared to existing methods.
• Demonstrated parameter sensitivity and scalability.

Conclusions:

• CDE++ effectively captures complex feature value couplings.
• The method enhances performance in downstream machine learning tasks.
• CDE++ offers a robust and scalable solution for categorical data representation.