阅读和尝试理解 TripleRE: Knowledge Graph Embeddings via triple Relation Vectors

文章来源

在open graph benchmark的比赛中,360的模型在数据集ogbl-wikikg2上拿了第一和第三，所以把论文拿过来看看

第一名的文章链接：
TripleRE: Knowledge Graph Embeddings via triple Relation Vectors

尝试理解和阅读文章

Introduction

在这里文中提到

Our work mainly lies in the optimization of the Translation distance model. One major theoretical issue that has dominated the field for many years concerns how to model the complex relation.

However, pairRE still only regards the relationship as the projection of the node. We believe that both pairRE and Rotate does not take account of the relationship can learn as the translation part of the node.

在文中，360的模型将relationship分解成了三个部分

We split the relationship into three parts.

• The projection part is the same as PairRE
• The translation part is learned by a separate parameter.
• 然后第三个部分好像没有明说(也可能是我没找到 QAQ)

不过在这个地方着重说明了TripleRe + NodePiece效果是最好的

Related Work

先不谈，都是前人的工作

Methodology

Loss Function

Score Function

Conclusions and Future Work

这篇文章主要就是证明了distance-based knowledge representation model也可以学习比较复杂的knowledge representation vectors

---

以后想到再补充