ON LEARNABILITY AND EXPERIENCE REPLAY METHODS FOR GRAPH INCREMENTAL LEARNING ON EVOLVING GRAPHS
Primary Area: transfer learning, meta learning, and lifelong 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: graph neural network, incremental learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Recent research has witnessed a surge in the exploration of Node-wise Graph Incremental Learning (NGIL) due to its substantial practical relevance. A central challenge in NGIL stems from the structural shifts induced by the inherent interdependence among vertices within graph data, adding complexity to the task of maintaining consistent model performance over time. Although several efforts have been made to devise incremental learning methods for NGIL, they have overlooked a fundamental question concerning the learnability of NGIL—whether there always exists a learning algorithm capable of consistently producing a model with a small error from the hypothesis. In this paper, we present the first theoretical study on the learnability of the NGIL problem with the statistical learning framework. Our analysis uncovers a critical insight: NGIL is not always learnable when structural shifts are uncontrolled. Additionally, in order to control structural shift, we leverage the idea of experience reply which selects a small set of representative data to replay with the new tasks, and propose a novel experience replay method, Structure-Evolution-Aware Experience Replay (SEA-ER). SEA-ER comprises a novel experience sample selection strategy founded on the topological awareness of GNNs and a novel replay objective utilizing importance re-weighting, which can effectively counteract catastrophic forgetting and mitigate the effect of structural shifts in NGIL. Comprehensive experiments validate our theoretical results and showcase the effectiveness of our newly proposed experience replay approach.
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: 2965
Loading