Skip to main content

Table 5 A list of hyperparameters optimised for each architecture type, and the domains over which they were optimised

From: Building attention and edge message passing neural networks for bioactivity and physical–chemical property prediction

HyperparameterSELU-MPNNAMPNNEMNN
Learn-rate{\(10^{ - 6} - 10^{ - 4}\)}{\(10^{ - 6} - 10^{ - 4}\)}{\(10^{ - 6} - 10^{ - 4}\)}
Message-size[10,16,25,40][10,16,25,40]NA
Message-passes[1–10][1–10][1–8]
Msg-hidden-dim[50,85,150][50,85,150][50,85,150]
Gather-width[30,45,70,100][30,45,70,100][30,45,70,100]
Gather-emb-hidden-dim[15,26,45,80][15,26,45,80][15, 26, 45]
Gather-att-hidden-dim[15,26,45,80][15,26,45,80][15, 26, 45]
Out-hidden-dim[360,450,560][360,450,560][360,450,560]
Out-dropout-p{0.0–0.1}{0.0–0.1}{0.0–0.1}
Out-layer-shrinkage{0.2–0.6}{0.2–0.6}{0.2–0.6}
Att-hidden-dimNA[50,85,150][50,85,150]
Edge-emb-hidden-dimNANA[60,105,180]
Edge-embedding-sizeNANA[30,50,80]
  1. Square brackets indicate discrete domains
  2. NA not applicable