Gated Graph Recursive Neural Networks for Molecular Property Prediction

Gated Graph Recursive Neural Networks for Molecular Property Prediction

Hiroyuki Shindo, Yuji Matsumoto

Gated Graph Recursive Neural Networks (GGRNet)

三个特点

1. 所有的projection matrix在不同的layer之间是shared
2. 每层layer都有input feature输入，作为skip connections
3. 根据距离进行message passing，而非adjacency matrix

考虑以下函数作为message-passing function

$$ p^t{v,w} = W_p \cdot [(x_v, h_v^t, x_w, h_w^t, x_N, d{v,w}] + bp\ q^t{v,w} = Wq \cdot [(x_v, h_v^t, x_w, h_w^t, x_N, d{v,w}] + bq\ m^{t+1}{v,w} = \sigma ( p^t{v,w}) \odot tanh(q^t{v,w})\ mv^{t+1} = \sum{w \ne v} m_{v,w}^{t+1}

$$

其中$$d_{v,w}$$是距离的倒数。