MaskDGNN: Self-Supervised Dynamic Graph Neural Networks with Activeness-aware Temporal Masking

Yiming He, Xiang Li, Zhongying Zhao, Haobing Liu, Peilan He, Yanwei YuIJCAI 2025 · OpenReview

TL;DR

This paper introduces MaskDGNN, a novel dynamic graph neural network that tackles two key challenges in modeling evolving networks: redundant information and distribution shifts. By employing an "activeness-aware temporal masking" strategy to focus on important connections and an "adaptive frequency enhancing" module to better handle changing patterns, MaskDGNN significantly improves performance on link prediction tasks.

What Problem Does This Paper Address?

Dynamic graphs, which represent evolving relationships over time (think social networks, traffic patterns, or e-commerce interactions), are crucial for understanding complex systems. While Graph Neural Networks (GNNs) are powerful for static graphs, adapting them to dynamic settings presents challenges. Existing dynamic GNNs often struggle with two main issues:

  1. Redundant Information: They can get bogged down by historical data, overemphasizing less active nodes and infrequent connections, which dilutes the signal and hinders performance.
  2. Distribution Shifting: The underlying patterns and relationships in dynamic graphs can change abruptly over time (e.g., market crashes affecting trading behavior). Models need to adapt to these shifts to maintain accuracy.

This paper aims to address these limitations by developing a dynamic GNN that can intelligently filter out noise and adapt to evolving graph structures.

Methods

MaskDGNN is built upon two novel components:

  1. Self-Supervised Activeness-Aware Temporal Masking Module: This module is designed to combat redundancy. It dynamically assesses the "activeness" of nodes at each time step. This activeness is a combination of:
  2. Node Dynamics Score: Measures how much a node's connections have changed between consecutive time steps.
  3. Node Significance Score: Evaluates a node's global importance within the current graph snapshot, similar to a PageRank score. By comparing these scores, the module selectively masks edges connecting less active nodes, thereby focusing the model's attention on the most critical interactions. The masking is done in a self-supervised manner, treating masked edges as positive samples to retain crucial link information.

  4. Adaptive Frequency Enhancing Graph Representation Learner: This module tackles the distribution shifting problem. It leverages Fourier transforms to analyze the frequency components of node embeddings. By adaptively enhancing key frequency components, the model can better capture intrinsic features that are resilient to temporal shifts. This is complemented by a sliding window strategy to aggregate long-term dependencies, making the model more robust to macro-level structural changes.

Datasets

The proposed MaskDGNN was evaluated on five real-world dynamic graph datasets:

  • Bitcoin-Alpha: A cryptocurrency trading network.
  • Bitcoin-OTC: Another cryptocurrency trading network.
  • UCI-Message: A communication network.
  • MOOC: A dataset related to online courses.
  • Wiki-Talk: A large-scale dynamic graph dataset of user interactions on a wiki platform, notable for its size (over 1 million nodes and 7 million edges).

Key Insights and Ideas

  • Focusing on "Activeness" is Key: The paper highlights that not all nodes and edges are equally important in dynamic graphs. Prioritizing connections between highly active nodes is crucial for effective learning and filtering out noise.
  • Frequency Domain for Temporal Shifts: Analyzing graph data in the frequency domain (using Fourier transforms) offers a novel way to capture and adapt to changing patterns that might be missed by purely time-domain approaches.
  • Self-Supervision for Robustness: Using masked edges as positive samples in a self-supervised loss helps the model learn to reconstruct and understand the underlying graph structure, even when parts are intentionally hidden.
  • Scalability Matters: The success on the large-scale Wiki-Talk dataset, where many baselines failed due to memory constraints, demonstrates the practical applicability of MaskDGNN.

Key Results

MaskDGNN demonstrated superior performance across all evaluated metrics (Accuracy, AUC, MRR, and Recall@10) on the five real-world datasets.

  • Overall Improvement: Achieved an average improvement of 7.07% in Accuracy and 13.87% in MRR compared to state-of-the-art methods.
  • Specific Dataset Wins: For example, on Bitcoin-Alpha, MaskDGNN boosted MRR by 35.86% and Recall@10 by 15.60%.
  • Scalability: Successfully handled the large Wiki-Talk dataset, outperforming baselines that encountered out-of-memory errors.
  • Ablation Studies: Removing either the activeness-aware masking or the frequency-enhancing layer led to significant performance drops, confirming the effectiveness of both proposed modules.

Limitations and Future Work

  • Hyperparameter Sensitivity: The paper notes that performance is sensitive to hyperparameters like the mask ratio, node dynamics score ratio, embedding dimension, and sliding window size. While they provide optimal ranges, careful tuning is required.
  • Computational Cost: While MaskDGNN is efficient, the sliding window mechanism and frequency analysis do add computational overhead, especially for very long sequences or extremely large graphs.
  • Future Directions: The authors suggest exploring more sophisticated frequency-domain techniques and investigating the model's applicability to other dynamic graph tasks beyond link prediction. Further research into more robust and adaptive masking strategies could also be beneficial.

Summary updated

Abstract

Integrating dynamics into graph neural networks (GNNs) provides deeper insights into the evolution of dynamic graphs, thereby enhancing the temporal representation in real-world dynamic network problems. Existing methods extracting critical information from dynamic graphs face two key challenges, either overlooking the negative impact of redundant information or struggling in addressing the distribution shifting issue in dynamic graphs. To address these challenges, we propose MaskDGNN, a novel dynamic GNN architecture that consists of two modules: First, self-supervised activeness-aware temporal masking mechanism selectively retains edges between highly active nodes while masking those with low activeness, effectively reducing redundancy. Second, adaptive frequency enhancing graph representation learner amplifies the frequency-domain features of nodes to capture intrinsic features under distribution shifting. Experiments on five real-world dynamic graph datasets demonstrate that MaskDGNN outperforms state-of-the-art methods, achieving an average improvement of 7.07% in accuracy and 13.87% in MRR for link prediction tasks.
BibTeX
@inproceedings{he2025maskdgnn,
  author = {Yiming He and Xiang Li and Zhongying Zhao and Haobing Liu and Peilan He and Yanwei Yu},
  title = {{MaskDGNN: Self-Supervised Dynamic Graph Neural Networks with Activeness-aware Temporal Masking}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=iEtXe70Bge},
}
Figure