Authors: William L. Hamilton (wleif@stanford.edu), Rex Ying (rexying@stanford.edu)
Attention! This is modified version of original codebase provided by authors of the paper.
I introduced next improvements:
- networkx is replaced with graph-tool* (allowed fast graph loading from binary file; faster dataset building for unsupervised training)
- train/test split is now happening on the fly for both supervised/unsupervised setups (no need to label nodes in advance)
- adjacency matrix is now being cached (speed-up setup before training)
- fixed issue with features matrix > 2Gb
- models are now stored as TF-saved_model (in the end of training for unsupervised, best model on f1-micro for supervised)
- inference is moved out
- two fully-connected layers were added after convolutions according to Pinterest paper ( Graph Convolutional Neural Networks for Web-Scale Recommender Systems by Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, Jure Leskovec ) Kudos to authors of original code!
*graph-tool installation:
https://git.skewed.de/count0/graph-tool/wikis/installation-instructions