Continual Knowledge Graph Link Prediction: Beyond Experience Replay

Han Wu, Jie Yin, Thang D Bui, Balakanapathy Rajaratnam

Keywords: Knowledge Graph Link Prediction, Continual Learning

Abstract:

Knowledge graphs (KGs) empower AI systems with essential inference capabilities as they increasingly integrate into life and industries. The dynamic nature of real-world KGs underscores the necessity for KG link prediction methods to possess continual learning capabilities.
However, the existing benchmark datasets primarily rely on sampling based methods, falling short of adequately evaluating models' abilities for continual KG link prediction. In this paper, we explicitly formulate the continual KG link prediction task and provide definitions for its two specific settings: class-incremental and expansive.
Two new benchmark datasets are established to provide valid benchmarking for fair evaluation of continual KG link prediction methods.
Furthermore, we propose BER, a novel approach based on experience replay and knowledge distillation to alleviate the catastrophic forgetting problem. Extensive experimental results demonstrate the datasets' effectiveness in providing a fair evaluation of continual learning ability and validate the efficacy of our proposed method. Codes can be found in supplementary material and will be released along with both datasets upon acceptance.

Primary Area: datasets and benchmarks

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.

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

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: 7309