[Release] DGL v0.4 Release (Heterogeneous Graph Update)

We are thrilled to announce the 0.4 release! This includes:

A heterogeneous graph is a graph whose nodes and edges are typed:

This is a companion discussion topic for the original entry at https://www.dgl.ai/release/2019/10/08/release.html

Thank you for the new release. Can’t wait to try it out.


Hi eveybody, was reading this paper on Graph Transformer Network (https://arxiv.org/pdf/1911.06455.pdf) for heterographs and it seems quite interesting. It learns a new graph structure that involves automatically identifying useful meta-paths and multi-hop connections for learning effective node representation on graphs in an end-to-end fashion unlike HAN and metapath2vec which involve manually selecting meta-paths by domain experts and thus might not be able to capture all meaningful relations for each problem. The learnt graph structures lead to more effective node representation resulting in state-of-the art performance, without any predefined meta-paths from domain knowledge. From the paper, it seems to outperform both metapath2vec and HAN. Interesting thing in the paper is how GAT somehow outperforms HAN demonstrating how the choice of meta-paths is important. Results in the HAN paper were different though. I think this would be a good addition to the DGL arsenal! What do you guys think? Code available at https://github.com/seongjunyun/Graph_Transformer_Networks

Nice, would you like to contribute? If you don’t have time, we could some time to implement this model too.

It seems not too complicated. I am going to try to implement it this weekend if I am free at that time. Let me know if you have already written some code (:

Hi Mika. I just took a look at the paper, their improvment is not based on GNN structure but maily about generating implicit meta-paths, you can use this api in the GNN part. For generating meta paths I think it’s effective and neat enough to use authors’ code.

Hi, Great to have support for heterogeneous graphs. I have been exploring DGL 4.0 for heterogeneous graphs.
Does DGL graph store support HeteroGraphs?? For a giant heterograph server storage is must.
Also, graph save and load functions are not extended to heterographs yet. I am struggling with having to upload my networkX graph everytime and then converting it to DGLHetero before running even a small experiment. How soon can we expect these updates?