A novel model for relation prediction in knowledge graphs exploiting semantic and structural feature integration

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Science

Background:

• Relation prediction is crucial for knowledge graph completion and knowledge representation.
• Existing methods often focus on either structural or semantic features, leading to incomplete knowledge representation.
• A unified approach to leverage both feature types is needed for improved relation prediction accuracy.

Purpose of the Study:

• To develop a novel model, RP-ISS, that integrates deep semantic and structural features for enhanced relation prediction.
• To address the limitations of existing methods that typically focus on a single feature type.
• To improve the accuracy and robustness of knowledge graph completion.

Main Methods:

• RP-ISS employs a two-part architecture: a RoBERTa module for semantic feature extraction and an edge-based relational message-passing network for structural information.
• A node embedding memory bank is utilized to mitigate computational burden during message passing.
• The model was evaluated on WN18RR, WN18, and FB15k-237 datasets.

Main Results:

• RP-ISS significantly surpasses all baseline methods across all evaluation metrics on the tested datasets.
• The model demonstrates superior performance in combining semantic and structural information for relation prediction.
• RP-ISS exhibits robust performance in graph inductive learning scenarios.

Conclusions:

• The proposed RP-ISS model effectively integrates deep semantic and structural features for superior relation prediction.
• RP-ISS offers a more comprehensive knowledge representation, overcoming limitations of previous single-feature approaches.
• The model's strong performance indicates its potential for advancing knowledge graph completion and related AI tasks.