The code to reproduce "EXPLORATION IN SEQUENTIAL RECOMMENDER SYSTEMS VIA GRAPH REPRESENTATIONS". Paper is prepared by D.Kiselev and I.Makarov in Artificial Intelligence Research Institure, Moscow, Russia.
The code is based on the original TGN implementation https://github.com/twitter-research/tgn. The main contribution lies in the file decision_module.py. Also, we have implemented the ope_loss_module.py with Replay counter-factual evaluation applied to the loss and modified train loop in train_ope_offline.py to run an online simulation.
To reproduce the paper run the notebook.