[Paper Review] DynWalks: Global Topology and Recent Changes Awareness Dynamic Network Embedding

Chengbin Hou, Han Zhang+2 more|arXiv (Cornell University)|Jul 27, 2019

Advanced Graph Neural Networks19 references4 citations

TL;DR

DynWalks proposes an online dynamic network embedding method that balances global topology preservation and recent structural changes through an adaptive node selection scheme and an efficient online learning framework. It achieves state-of-the-art performance in graph reconstruction and link prediction while significantly reducing wall-clock time compared to existing methods.

ABSTRACT

Learning topological representation of a network in dynamic environments has recently attracted considerable attention due to the time-evolving nature of many real-world networks i.e. nodes/links might be added/removed as time goes on. Dynamic network embedding aims to learn low dimensional embeddings for unseen and seen nodes by using any currently available snapshots of a dynamic network. For seen nodes, the existing methods either treat them equally important or focus on the $k$ most affected nodes at each time step. However, the former solution is time-consuming, and the later solution that relies on incoming changes may lose the global topology---an important feature for downstream tasks. To address these challenges, we propose a dynamic network embedding method called DynWalks, which includes two key components: 1) An online network embedding framework that can dynamically and efficiently learn embeddings based on the selected nodes; 2) A novel online node selecting scheme that offers the flexible choices to balance global topology and recent changes, as well as to fulfill the real-time constraint if needed. The empirical studies on six real-world dynamic networks under three different slicing ways show that DynWalks significantly outperforms the state-of-the-art methods in graph reconstruction tasks, and obtains comparable results in link prediction tasks. Furthermore, the wall-clock time and complexity analysis demonstrate its excellent time and space efficiency. The source code of DynWalks is available at https://github.com/houchengbin/DynWalks

Motivation & Objective

• To address the high computational complexity and poor global topology preservation in existing dynamic network embedding methods.
• To enable real-time learning for large-scale dynamic networks under strict time constraints.
• To develop a flexible node selection strategy that balances attention between global network structure and local changes.
• To maintain embedding quality across evolving snapshots without full retraining.
• To achieve high efficiency and scalability for online deployment in real-world dynamic networks.

Proposed method

• DynWalks employs an online network embedding framework based on modified DeepWalk that updates embeddings incrementally using only selected nodes at each time step.
• It introduces a novel node selection scheme that prioritizes nodes based on their impact on global topology and recent changes, controlled by hyperparameters α and β.
• The method uses embeddings from the previous time step (t−1) as initialization for the current step (t), improving convergence and stability.
• The framework dynamically selects a subset of nodes—typically a fraction α of the network—based on topological sensitivity and change impact.
• It balances global structure awareness (via β) and responsiveness to recent changes (via α), allowing trade-offs between accuracy and speed.
• The approach avoids full retraining by focusing only on affected nodes and leveraging incremental optimization.

Experimental results

Research questions

• RQ1 How can dynamic network embedding be made efficient enough to meet real-time constraints in large-scale evolving networks?
• RQ2 To what extent can a dynamic embedding method preserve the global topology of a network while responding to local changes?
• RQ3 Can a selective node update strategy outperform full-network retraining in terms of both accuracy and efficiency?
• RQ4 How does the balance between global topology awareness and responsiveness to recent changes affect downstream task performance?
• RQ5 What is the impact of using prior embeddings as initialization on convergence speed and stability?

Key findings

• DynWalks significantly outperforms state-of-the-art methods in graph reconstruction tasks across six real-world dynamic networks under three slicing strategies.
• It achieves comparable performance to SOTA in link prediction tasks, with gradual improvement as the network grows over time.
• The method reduces wall-clock time by up to 90% compared to BCGD and DynGEM, with online stage time dropping to under 4 seconds per step on large networks.
• Sensitivity analysis shows that performance saturates or slightly declines when α exceeds 0.6, indicating an optimal balance point for node selection.
• The use of t−1 embeddings as initialization contributes to stable performance over time, especially in graph reconstruction tasks.
• The method remains scalable and efficient even over 100 time steps, with minimal degradation in performance and strong time efficiency.