Bilinear Scoring Function Search for Knowledge Graph Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Knowledge graph (KG) embeddings and scoring functions are crucial for downstream tasks.
• Existing human-designed scoring functions struggle with complex, inferred relation patterns.
• There's a need for adaptive scoring functions tailored to specific KG tasks.

Purpose of the Study:

• To automatically search for and discover optimal bilinear scoring functions for KG learning using automated machine learning (AutoML).
• To develop advanced AutoML techniques (AutoBLM and AutoBLM+) that incorporate domain-specific information for KG tasks.
• To evaluate the effectiveness of the discovered scoring functions on various KG benchmarks.

Main Methods:

• Defined a search space for bilinear scoring functions by analyzing existing methods.
• Proposed AutoBLM, a progressive AutoML algorithm, and AutoBLM+, an enhanced evolutionary algorithm.
• Accelerated algorithms using filter and predictor mechanisms to handle domain-specific properties.
• Conducted extensive experiments on KG completion, multi-hop query, and entity classification tasks.

Main Results:

• Discovered novel, KG-dependent scoring functions that outperform existing methods.
• AutoBLM+ demonstrated superior performance compared to AutoBLM due to its evolutionary approach.
• The searched scoring functions showed significant improvements across diverse KG learning benchmarks.

Conclusions:

• AutoML is effective for discovering high-performing, task-specific KG scoring functions.
• The AutoBLM+ evolutionary approach offers more flexible exploration of scoring function structures.
• This work advances KG learning by automating the design of crucial scoring function components.