DAPath: Distance-aware knowledge graph reasoning based on deep reinforcement learning.

Area of Science:

• Artificial Intelligence
• Computer Science

Background:

• Knowledge graph reasoning (KGR) seeks to infer missing information in incomplete knowledge graphs (KGs).
• Existing methods often overlook the varying importance of different positions (vertices) within the graph during reasoning.
• Standard KG embeddings may not fully capture nuanced information from an entity's local neighborhood.

Purpose of the Study:

• To develop a novel reinforcement learning framework for KG reasoning that incorporates distance-aware rewards.
• To enhance the representation of entities by integrating graph self-attention (GSA) for richer local context.
• To improve path memory in reasoning by combining GSA with GRU for relation sequence awareness.

Main Methods:

• A reinforcement learning framework with a proposed distance-aware reward mechanism to assign differential rewards to graph vertices.
• Integration of a graph self-attention (GSA) mechanism to capture comprehensive entity information from neighboring entities and relations.
• Combination of GSA with Gated Recurrent Unit (GRU) to enable the model to remember relational paths.

Main Results:

• The proposed method enables one-pass agent training, eliminating the need for complex pre-training or fine-tuning processes.
• Experimental validation demonstrates the effectiveness of the approach in knowledge graph reasoning.
• The model successfully mines more balanced reasoning paths for various relations within the KG.

Conclusions:

• The novel approach effectively addresses limitations in prior knowledge graph reasoning techniques.
• The integration of distance-aware rewards and GSA-GRU significantly improves the quality and balance of discovered reasoning paths.
• This method offers a more efficient and effective solution for reasoning over incomplete knowledge graphs.