Core Concepts
The core message of this paper is to propose a reinforcement learning-based framework, called Ada-DyGNN, that can adaptively determine which nodes should be updated when new connections are established in dynamic graphs. This is in contrast to previous approaches that blindly update the embeddings of all neighboring nodes, which can lead to performance degradation when the new connections contain noisy or outdated information.
Abstract
The paper introduces Ada-DyGNN, a robust knowledge adaptation framework for dynamic graph neural networks. The key highlights are:
Existing dynamic graph neural network models often blindly update the embeddings of all neighboring nodes when new connections are established. This can be problematic when the new connections contain noisy or outdated information.
Ada-DyGNN addresses this issue by using reinforcement learning to adaptively determine which nodes should be updated. It formulates the node update selection as a sequence decision problem and employs reinforcement learning to optimize the policy network.
The paper proposes a new reward function that encourages the stability of local structures, defined based on neighbor similarity. This helps Ada-DyGNN learn robust node representations by selectively propagating knowledge.
Extensive experiments on three benchmark datasets demonstrate the effectiveness of Ada-DyGNN, especially in terms of its robustness to noise compared to existing methods.