Triplet-aware graph neural networks for factorized multi-modal knowledge graph entity alignment.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Multi-Modal Entity Alignment (MMEA) is crucial for knowledge graph fusion, enabling the integration of disparate multi-modal knowledge graphs (MMKGs).
• Existing methods often focus on attribute aggregation, overlooking the interplay between multi-modal attributes and graph structure.
• This limitation hinders the effective integration and alignment of entities within MMKGs.

Purpose of the Study:

• To propose an innovative approach, TriFac, for effective Multi-Modal Entity Alignment.
• To address the limitations of current methods by incorporating both structural and multi-modal attribute information.
• To enhance the performance of knowledge graph fusion through improved entity alignment.

Main Methods:

• Developed TriFac, a two-stage MMKG factorization approach utilizing embedding refinement.
• Employed triplet-aware graph neural networks for aggregating multi-relational features.
• Implemented multi-modal fusion techniques to integrate diverse features and introduced novel metrics for evaluating factorization performance.

Main Results:

• Empirical results demonstrate the superior effectiveness of the TriFac approach.
• TriFac significantly outperforms previous state-of-the-art models on MMEA tasks.
• The model showed strong performance on two MMEA datasets and a specialized power system dataset.

Conclusions:

• TriFac provides an effective solution for Multi-Modal Entity Alignment by leveraging embedding refinement and graph factorization.
• The proposed method successfully integrates multi-modal attributes and graph structure, outperforming existing approaches.
• This work advances the field of knowledge graph fusion and entity alignment, particularly for complex multi-modal data.