i-Align: an interpretable knowledge graph alignment model

Abstract Knowledge graphs (KGs) are becoming essential resources for many downstream applications. However, their incompleteness may limit potential. Thus, continuous curation is needed to mitigate this problem. One of the strategies address problem KG alignment, i.e., forming a more complete by merging two or KGs. This paper proposes i-Align, an interpretable alignment model. Unlike existing models, i-Align provides explanation each prediction while maintaining high performance. Experts can use check correctness prediction. quality be maintained during process (e.g., KGs). To end, novel Transformer-based Graph Encoder (Trans-GE) proposed as key component aggregating information from entities’ neighbors (structures). Trans-GE uses Edge-gated Attention that combines adjacency matrix and self-attention learn gating mechanism control aggregation neighboring entities. It also historical embeddings , allowing trained over mini-batches, smaller sub-graphs, scalability issue when encoding large KG. Another Transformer encoder attributes. way, generate explanations in form set most influential attributes/neighbors based on attention weights. Extensive experiments conducted show power i-Align. The include several aspects, such model’s effectiveness aligning KGs, generated explanations, its practicality results these aspects.