Drug-Drug Adverse Effect Prediction with Graph Co-Attention

Drug-Drug Adverse Effect Prediction with Graph Co-Attention

3 Architecture

3.1 Inputs

有两类输入

1. 作为binary classification，输入的时候就考虑到了label
2. 作为multi-label classification，输入的时候只有drug pair，输出有多个label

3.2 Message Passing

正常操作

3.3 Co-attention

$$ a{ij}^t = softmax_j \Big( \langle W_k^{t} h_i^{t-1,\, d_x}, W_k^{t} h_i^{t-1, \, d_y} \rangle \Big)\ n_t^{t\, d(x)} = \sum{\forall j \in dy} a{ij}^t \cdot W_v^t h_j^{t-1, \, d_y}

$$

然后这个学到的co-attention layer会一起用来update

$$ h_i^{t,\, d_x} = LayerNorm(h_i^{t-1,\, d_x} + m_i^{t-1,\, d_x} + n_i^{t-1,\, d_x})

$$