Paper总结 - RotatE: Knowledge Graph Embedding by Relational Rotation in Complex Space
这篇文章发表在2019年的ICLR上
主要思想:将 \(h\),\(r\), \(t\) 用复数来表示,即 \(h\), \(t\), \(r\) \(\in\) \(C\)。
首先,任何复数 \(Z\) 可表示成: \(Z = r(\cos \theta+i \sin \theta) = r \cos \theta + ir \sin \theta\)
在RotatE中,约束 \(|r| = 1\),这样 \(r\) 可看作是一种旋转,譬如
\(h \circ r=h(\cos \theta+i \sin \theta)=t^{\prime}\), 这里的 \(\circ\) 是 element-wide product。
distance function:
\(d_{r}(h, t)=\left\|h \circ r-1\right\|\)
Loss function:
\(L=-\log \sigma\left(\gamma-d_{r}(\mathbf{h}, \mathbf{t})\right)-\sum_{i=1}^{n} \frac{1}{k} \log \sigma\left(d_{r}\left(\mathbf{h}_{i}^{\prime}, \mathbf{t}_{i}^{\prime}\right)-\gamma\right)\)
其中 \(d_{r}(\mathbf{h}, \mathbf{t})\) 是 negative sample。
RotatE 可以很好的 model 几乎所有关系:
Three types of relation:
- symmetric / antisymmetry eg. \(\left(h_{1}, r, h_{2}\right) \Leftrightarrow\left(h_{2}, r, h_{1}\right)\)
- inversion e.g. \(\left(h_{1}, r_{1}, h_{2}\right) \Leftrightarrow\left(h_{2}, r_{2}, h_{1}\right)\)
- composition e.g. \(\left(h_{1}, r_{1}, h_{2}\right) \wedge\left(h_{2}, r_{2}, h_{3}\right) \Rightarrow\left(h_{1}, r_{3}, h_{3}\right)\)
最后除了普通的Link prediction实验外,还进行了contry dataset的实验。