ConvR
Adaptive Convolution for Multi-Relational Learning
2019-6
1 Introduction
Learning with multi-relational data plays a pivotal role in many application domains, ranging from social networks or recommender systems to large-scale knowledge bases (KBs)
ConvR的思想是从relation中构造filter,然后卷积于subject embedding,最后投影,与object embedding做点积。
这样的做法就导致了ConvR的另一个优势,减少了参数的数量。
3 Adaptive Convolution on Multi-relational Data
对于三元组\((s, r, o)\),首先将\(e_s\) reshape为2D矩阵, \[ e_s\in R^{d_e}\Rightarrow S\in R^{d_e^h,\ d_e^w} \] 对于关系\(r\),先分割为\(c\)段: \[ r^{(1)},\cdots,r^{(c)} \] 然后每个\(r^{(l)}\) reshape为2D的矩阵作为filter \[ R^{l}\in R^{h,\ w} \] 对于\(c\)个filter,在\(S\)上卷积,得到\(c\)个feature map。
将\(c\)个feature map先展开为一维,然后stack到一起,得到单向量\(e_c\)。
最后过一个全连接层,和尾结点计算点积 \[ \psi(s,r,o)=f(We_c+b)e_o \] 训练方式与ConvE保持一致。
比起ConvE的好处就是结果更好,参数更少,空间复杂度降低。
ConvR使用三个dropout防过拟合:
- 在reshape subject representation时
- 在卷积得到feature map之后
- 在经过全连接之后
4 Experiments
使用了四个数据集:
- FB15k
- WN18
- FB15K-237
- WN18RR
实现的超参: