Propositionalization and embeddings: two sides of the same coin.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Data preprocessing, including transformation, is crucial for machine learning pipelines.
• Relational learning often requires fusing diverse data types into a single tabular format.
• Propositionalization and embedding are distinct techniques for this data transformation.

Purpose of the Study:

• To provide a unifying framework for understanding propositionalization and embedding techniques in relational learning.
• To introduce a novel methodology combining both approaches for complex data transformation tasks.
• To present efficient implementations and evaluate their performance on relational problems.

Main Methods:

• Developed a unifying framework defining propositionalization and embedding as variants of a complex data transformation task.
• Proposed a unifying methodology integrating propositionalization and embedding.
• Implemented two efficient approaches: instance-based PropDRM and feature-based PropStar.

Main Results:

• The unifying framework clarifies similarities and differences between propositionalization and embedding.
• The combined methodology leverages the strengths of both techniques.
• PropDRM and PropStar demonstrated superior performance compared to existing relational learners on several problems.
• The new algorithms successfully tackled larger-scale relational learning tasks.

Conclusions:

• The unified framework enhances the understanding of data transformation techniques in relational learning.
• The novel methodology offers an effective approach for complex data transformation and learning.
• Efficient implementations like PropDRM and PropStar provide practical solutions for advanced relational learning challenges.