This repository contains a Pytorch Geometric implementation of the Gravity-Inspired Graph Autoencoders for Directed Link Prediction paper.
This implementation has been referenced in the original implementation, and it has been reviewed by one of its authors.
With respect to the original implementation, it has the following limitations:
And additions:
See below for details on each of these additions.
The general setting contains a validation set that we use to perform early stopping w.r.t. a validation loss obtained by summing the AUC and AP scores on the validation set.
Instead of treating lambda as an hyperparameter, we enable training for it.
The "biased_rev" tasks have a train set which is different from that of the original implementation ("biased"). It consists of all positive edges (as defined in the "biased" task) and all their reverses as negatives. This produces a training set which is statistically more similar to the test set and thus improves performance on many models.
Sometimes the recon_loss is more numerically stable than the BCELossWithLogits, and thus we preferred it.
Please use the provided environment.yml file to instantiate a conda environment gravity_gae_env compatible with this code. Then open the main.py script, set the parameters within the "Please set the parameters below" section and run it e.g. in VSCode.
Salha, G., Limnios, S., Hennequin, R., Tran, V., & Vazirgiannis, M. (2019). Gravity-Inspired Graph Autoencoders for Directed Link Prediction. In ACM International Conference on Information and Knowledge Management (CIKM).