CoMemNet: A Dual-Branch Continual Learning Framework for Traffic Prediction
Researchers have introduced CoMemNet, a continual learning framework with dual branches designed for traffic prediction, which tackles the issue of catastrophic forgetting in streaming networks. The Online branch, known for its rapid convergence, focuses on making primary predictions, whereas the Target branch, updated using momentum, employs Wasserstein Distance features to develop a Dynamic Contrastive Sampler (DC Sampler). This sampler identifies nodes that exhibit notable dynamic changes for training, helping to reduce forgetting. Additionally, the framework combines non-topological space modeling with temporal learning to effectively handle non-Euclidean graphs.
Key facts
- CoMemNet is a dual-branch continual learning framework for traffic prediction.
- It addresses catastrophic forgetting in streaming traffic networks.
- The Online branch handles primary prediction tasks.
- The Target branch uses Wasserstein Distance features for a Dynamic Contrastive Sampler.
- The DC Sampler selects nodes with significant dynamic feature changes.
- The framework integrates non-topological space modeling with temporal learning.
- It targets non-Euclidean graphs in traffic networks.
- The method is proposed for continuously expanding and evolving traffic patterns.
Entities
—