Double Equivariance for Inductive Link Prediction for Both New Nodes and New Relation Types

Jianfei Gao, Yangze Zhou, Jincheng Zhou, Bruno Ribeiro

Primary Area: learning on graphs and other geometries & topologies

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: Link Prediction, Double Permutation-Equivariance, Discrete Attributed Multigraph, Knowledge Graph, GNN

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

TL;DR: This work describes the doubly inductive link prediction task (over new nodes and new relation types in test) and introduce the associated double equivariant models

Abstract: The task of inductive link prediction in discrete attributed multigraphs (e.g., knowledge graphs, multilayer networks, heterogeneous networks, etc.) generally focuses on test predictions with solely new nodes but not both new nodes and new relation types. In this work, we formally define the task of predicting (completely) new nodes and new relation types in test as a doubly inductive link prediction task and introduce a theoretical framework for the solution. We start by defining the concept of double permutation-equivariant representations that are equivariant to permutations of both node identities and edge relation types. We then propose a general blueprint to design neural architectures that impose a structural representation of relations that can inductively generalize from training nodes and relations to arbitrarily new test nodes and relations without the need for adaptation, side information, or retraining. We also introduce the concept of distributionally double equivariant positional embeddings designed to perform the same task. Finally, we empirically demonstrate the capability of the two proposed models on a set of novel real-world benchmarks, showcasing relative performance gains of up to 41.40% on predicting new relations types compared to baselines.

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: 8054