Yesterday, I found attractive article in arXiv.
Attentional Multilabel Learning over Graphs
A message passing approach
Click to access 1804.00293.pdf
I have not fully understood the concept but interesting.
The author proposes Graph Attentional Multi-Label learning (GAML). GAML handle Graph structured data as input. You know, molecules are represented as graph in chemoinformatics. So this method will be useful for chemoinformatics area.
They use basic graph neural network structure and use Highway for model update. (*Highway network is very interesting for me!)
Also they use attention to connect every input node and label node. Table3 shows GAML works very well than other methods for example SVM, HWN(highway), GRU etc.
I think attention method is very informative because the method can draw attention matrix. The author shows attention matrix between label and input(graph structure of molecules). And the heat map shows attention of each node(atom).
The code is not opened but there are several implementation of Graph attention networks and attentions.
I would like to try to implement GAML near the future.