Towards Robust Dynamic Network Embedding

Chengbin Hou, Ke TangIJCAI 2021 · DOI

Abstract

Dynamic Network Embedding (DNE) is essential for modeling evolving systems, yet existing methods are highly sensitive to how continuous data is discretized into snapshots. This paper identifies that data slicing strategies—fixed time intervals versus fixed edge counts—significantly impact model performance. We propose an ensemble-based framework that enhances diversity among base learners to achieve more stable embeddings across varying degrees of network change.

The Challenge of Network Slicing

Dynamic networks, such as social interactions or communication logs, are typically represented as a stream of timed edges. To apply DNE algorithms, researchers must partition this stream into a sequence of snapshots. Two common strategies are employed: Fixed Edge Slicing, where each snapshot contains a set number of new edges, and Fixed Time Slicing, where snapshots are defined by temporal windows (e.g., daily).

Illustration of raw edge data and different snapshot slicing strategies (G1 and G2).

The core problem identified is that a model's effectiveness is often tied to the "degree of change"—the average number of edge modifications between snapshots. A model optimized for steady, fixed-edge changes may perform poorly when faced with the erratic bursts of activity common in fixed-time slicing. Current DNE methods lack the robustness to handle these fluctuations, making their performance unpredictable in real-world scenarios where the rate of change is non-uniform.

Proposed Methodology: Robust Ensemble DNE

To address this instability, we propose a framework centered on ensemble learning. Rather than relying on a single embedding model, our approach utilizes multiple "base learners" to capture a broader spectrum of network dynamics.

Diversity Enhancement Strategy

The key innovation is a diversity enhancement strategy. We employ an incremental version of the Skip-Gram model as the base learner. By forcing different learners to prioritize different structural properties—ranging from local neighborhoods to global network topology—the ensemble becomes less sensitive to the specific discretization of the timeline. This redundancy ensures that if one learner is adversely affected by a sudden spike in edge activity, others can maintain the integrity of the overall embedding.

Evaluation and Results

The framework was benchmarked against six state-of-the-art DNE methods using five real-world datasets and various synthetic slicing configurations.

Key Findings

  1. Sensitivity to Slicing: Empirical results confirm that existing DNE methods are not robust. A method's rank in performance can shift dramatically depending on whether fixed-edge or fixed-time slicing is used.
  2. Ensemble Performance: The proposed ensemble method, SG-EDNE, demonstrated superior robustness. It consistently outperformed or matched single-model approaches across downstream tasks such as node classification and link prediction, regardless of the slicing strategy.
  3. Stability: The use of diverse views within the ensemble effectively smoothed out performance dips caused by high-variance network changes.

Conclusion and Future Directions

This research highlights that data preprocessing (slicing) is a critical variable in DNE performance, not merely a trivial step. While our ensemble approach provides a significant step toward robustness, handling the extreme variance in fixed-time slicing remains a challenging open problem. Future work will focus on optimizing base learner cooperation and applying these robust embeddings to high-stakes domains like real-time anomaly detection in industrial networks.

Summary updated

BibTeX
@inproceedings{hou2021towards,
  author = {Chengbin Hou and Ke Tang},
  title = {{Towards Robust Dynamic Network Embedding}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/676},
}
Figure