Robust Knowledge Adaptation for Dynamic Graph Neural Networks

TL;DR

Ada-DyGNN addresses the challenge of robust knowledge propagation in dynamic graphs where new links can introduce noisy information. It integrates a time-aware attentional aggregating module with a reinforced neighbor selection module, trained via reinforcement learning to decide which neighbors to update, guided by a reward that preserves local structural stability. This approach treats neighbor updates as a sequential decision problem and demonstrates state-of-the-art performance on three benchmarks, with strong robustness to varying noise levels. The work significantly advances dynamic graph learning by enabling selective information propagation that mitigates noise and outdated links, with practical implications for real-world temporal networks.

Abstract

Graph structured data often possess dynamic characters in nature. Recent years have witnessed the increasing attentions paid to dynamic graph neural networks for modelling graph data. However, almost all existing approaches operate under the assumption that, upon the establishment of a new link, the embeddings of the neighboring nodes should undergo updates to learn temporal dynamics. Nevertheless, these approaches face the following limitation: If the node introduced by a new connection contains noisy information, propagating its knowledge to other nodes becomes unreliable and may even lead to the collapse of the model. In this paper, we propose Ada-DyGNN: a robust knowledge Adaptation framework via reinforcement learning for Dynamic Graph Neural Networks. In contrast to previous approaches, which update the embeddings of the neighbor nodes immediately after adding a new link, Ada-DyGNN adaptively determines which nodes should be updated. Considering that the decision to update the embedding of one neighbor node can significantly impact other neighbor nodes, we conceptualize the node update selection as a sequence decision problem and employ reinforcement learning to address it effectively. By this means, we can adaptively propagate knowledge to other nodes for learning robust node embedding representations. To the best of our knowledge, our approach constitutes the first attempt to explore robust knowledge adaptation via reinforcement learning specifically tailored for dynamic graph neural networks. Extensive experiments on three benchmark datasets demonstrate that Ada-DyGNN achieves the state-of-the-art performance. In addition, we conduct experiments by introducing different degrees of noise into the dataset, quantitatively and qualitatively illustrating the robustness of Ada-DyGNN.

Figures (6)

• Figure 1: An illustration of noise and outdated links. A new link is built between $v_1$ and $v_2$ at timestamp $t_5$. The link between $v_1$ and $v_3$ is outdated (assume $t_0$ is too much earlier than $t_5$). $v_7$ is the node that contains noisy information.
• Figure 2: The overall architecture of our Ada-DyGNN method. When a new link between $v_1$ and $v_2$ is added, Ada-DyGNN performs robust knowledge propagation by the following steps: 1) The interaction message $\mathbf{m}_1$ and $\mathbf{m}_2$ are aggregated from $v_1$'s neighborhood $\{v_3, v_4, v_5\}$ and $v_2$'s neighborhood $\{v_6, v_7\}$, respectively; 2) The concatenated interaction message $\mathbf{m}$ propagates the information to $\{v_1,\dots,v_7\}$ to obtain the intermediate states $\{\mathbf{h}_1,\dots,\mathbf{h}_7\}$ based on the time-aware attentional aggregating module, after the connection between $v_1$ and $v_2$ is established; 3) The states are constructed by the node embeddings and the intermediate states, and are sent to the policy network as the inputs; 4) The reinforced neighbor selection module outputs actions to update or retain the embedding of each node; 5) The influenced nodes are updated based on the intermediate embeddings and a MLP. The reward is used to optimize the policy network.
• Figure 3: The training of our policy network.
• Figure 4: Actions taken by our neighbor selection module under different levels of noise.
• Figure 5: The reward curves of our method on the UCI, Wikipedia, and Reddit datasets, respectively.