TL;DR
This paper introduces the Online Continual Graph Learning (OCGL) setting, formalizing node-level learning on evolving graphs under strict resource constraints, and provides a benchmark and baseline methods for this new task.
Contribution
It formalizes the OCGL setting, creates a comprehensive benchmark with datasets and strategies, and proposes a minimalistic baseline for efficient online graph learning.
Findings
The benchmark enables standardized evaluation of OCGL methods.
The baseline achieves strong empirical performance with high efficiency.
OCGL effectively addresses resource constraints in online graph learning.
Abstract
Continual Learning (CL) aims to incrementally acquire new knowledge while mitigating catastrophic forgetting. Within this setting, Online Continual Learning (OCL) focuses on updating models promptly and incrementally from single or small batches of observations from a data stream. Extending OCL to graph-structured data is crucial, as many real-world networks evolve over time and require timely, online predictions. However, existing continual or streaming graph learning methods typically assume access to entire graph snapshots or multiple passes over tasks, violating the efficiency constraints of the online setting. To address this gap, we introduce the Online Continual Graph Learning (OCGL) setting, which formalizes node-level continual learning on evolving graphs under strict memory and computational budgets. OCGL defines how a model incrementally processes a stream of node-level…
Peer Reviews
Decision·Submitted to ICLR 2025
1. One of the paper's main strengths is the formal introduction of the Online Continual Graph Learning (OCGL) framework. 2. The authors develop a benchmarking environment specifically for OCGL, including multiple datasets and evaluations of various continual learning methods. 3. The experimental setup and the detailed analysis provided in the paper are thorough and well-constructed.
1. The proposed problem is novel, however, the detailed appliable scenario for such OCGL framework should be further explained, especially on graph data. 2. The baselines chosen in this paper are all Continual learning methods. More methods for the online learning setting should be included. 3. Also, as a benchmark paper, it would be beneficial to introduce more new datasets. 4. The contribution of this paper seems limited to me. The authors introduced a new problem setting OCGL for graph learn
1. Online continual graph learning has not been fully explored, and this work makes some contribution in this direction. 2. Compared to existing continual graph learning works, this work adopts a more practical hyperparameter selection strategy that only use a few tasks.
1. The main weakness is the inconsistence between the proposed setting and the actual experiments. Although the paper described an online learning setting, but the task construction in experiments is still same as the continual graph learning setting with task boundaries. As mentioned in the paper, 'the graph grow with nodes from two new classes at a time', then the incremental manner is same as a normal class incremental learning instead of an online learning setting. I would recommend that the
1. The introduction of the Online Continual Graph Learning (OCGL) framework extends continual learning to dynamic, graph-structured data. 2. The paper provides a thorough evaluation of multiple continual learning methods, adapting them for online graph learning. 3. The proposed neighborhood sampling strategy effectively addresses the computational and memory challenges of multi-hop neighborhood aggregation in GNNs.
1. The benchmarks focus mainly on node classification tasks, and extending the framework to more diverse graph-based applications (e.g., edge prediction, link prediction) could strengthen the paper's contributions. 2. The paper primarily compares traditional continual learning methods adapted for the Online Continual Graph Learning (OCGL) framework. It does not include comparisons with more recent state-of-the-art continual graph learning methods proposed in the recent three years, such as MSCG
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
