Training graph backward compatible embeddings

Daniil Bondarev, Dmitrii Kiselev

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Contrastive learning, embeddings, supervised, graph neural nework

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

In this paper we study the problem of training backward compatible embeddings.
We suggest framework to train neural networks end-to-end using contrastive learning. Applying contrastive learning allow us to improve exciting method, BCAligner, for solving backward compatibility problem and obtain significant metrics improvement. We consider graph neural networks (GCN, GAT) for our experiments and use open-source graph datasets (Cora, CiteSeer) to compare methods. Our method outperform BC-Aligner by 6%/5% accuracy for Cora/CiteSeer
datasets correspondingly.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9060