Revealing an Overlooked Challenge in Class-Incremental Graph Learning

TMLR Paper1966 Authors

20 Dec 2023 (modified: 06 May 2024)Decision pending for TMLREveryoneRevisionsBibTeX
Abstract: Graph Neural Networks (GNNs), which effectively learn from static graph-structured data, become ineffective when directly applied to streaming data in a continual learning (CL) scenario. In CL, historical data are not available during the current stage due to a number of reasons, such as limited storage, GDPR1 data retention policy, to name a few. A few recent works study this problem, however, they overlook the uniqueness of continual graph learning (CGL), compared to well-studied continual image classification: the unavailability of previous training data further poses challenges to inference in CGL, in additional to the well-known catastrophic forgetting problem. While existing works make a strong assumption that full access of historical data is unavailable during training but provided during inference, which potentially contradicts the continual learning paradigm Van de Ven & Tolias (2019), we study continual graph learning without this strong and contradictory assumption. In this case, without being re-inserted into previous training graphs for inference, streaming test nodes are often more sparsely connected, which makes the inference more difficult due to insufficient neighborhood information. In this work, we propose ReplayGNN (ReGNN) to jointly solve the above two challenges without memory buffers: catastrophic forgetting and poor neighbor information during inference. Extensive experiments demonstrate the effectiveness of our model over baseline models and its effectiveness in different cases with different levels of neighbor information available.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Manzil_Zaheer1
Submission Number: 1966