Online Lifelong Learning on Streaming Graphs

Continual Learning (CL) aims to learn new tasks incrementally while mitigating catastrophic forgetting. Online Continual Learning (OCL) narrows this focus to learning efficiently from a continuous, dynamically shifting data stream. Although recent studies have explored Continual Learning on graphs using Graph Neural Networks (GNNs), few address the unique challenges of streaming settings, where real-world graphs often evolve over time and require timely (online) predictions. However, current approaches to node classification are not well aligned with standard OCL frameworks, partly due to the lack of a clear definition of online Continual Learning for graphs. In this thesis, we propose a comprehensive framework for online Continual Learning on graphs, emphasizing efficient batch processing that accounts for graph topology. We introduce benchmark datasets for online continual graph learning and report the performance of various CL methods adapted to this framework. We further tackle the memory challenges posed by multi-hop neighborhood aggregation in GNNs, as it can require access to an ever-expanding graph, which leads to prohibitive costs in this setting. To address this, we introduce a sampling strategy to control complexity for efficient online learning. Finally, we investigate the potential of combining an untrained feature extractor with a linear classifier to mitigate forgetting.