Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Changepoint Detection in Highly-Attributed Dynamic Graphs

Emiliano Penaloza, Nathaniel Stevens

TL;DR

The work tackles detecting changepoints in dynamic graphs with high-dimensional node attributes by tracking community structure through modularity and estimating it with a GNN-based Deep Modularity Network (DMoN). It introduces a square root collapse optimizer to stabilize modularity optimization and couples Phase I/II statistical process monitoring with EWMA charts to detect changes in the modularity signal. Through synthetic experiments on DCSBM-based graphs and a real-world Twitter dataset from the #Iran network, the approach detects structural and attribute changes quickly with low false alarms. The framework provides a principled, attribute-aware method for monitoring evolving networks and is accompanied by data and code for reproducibility.

Abstract

Detecting anomalous behavior in dynamic networks remains a constant challenge. This problem is further exacerbated when the underlying topology of these networks is affected by individual highly-dimensional node attributes. We address this issue by tracking a network's modularity as a proxy of its community structure. We leverage Graph Neural Networks (GNNs) to estimate each snapshot's modularity. GNNs can account for both network structure and high-dimensional node attributes, providing a comprehensive approach for estimating network statistics. Our method is validated through simulations that demonstrate its ability to detect changes in highly-attributed networks by analyzing shifts in modularity. Moreover, we find our method is able to detect a real-world event within the \#Iran Twitter reply network, where each node has high-dimensional textual attributes.

Changepoint Detection in Highly-Attributed Dynamic Graphs

TL;DR

The work tackles detecting changepoints in dynamic graphs with high-dimensional node attributes by tracking community structure through modularity and estimating it with a GNN-based Deep Modularity Network (DMoN). It introduces a square root collapse optimizer to stabilize modularity optimization and couples Phase I/II statistical process monitoring with EWMA charts to detect changes in the modularity signal. Through synthetic experiments on DCSBM-based graphs and a real-world Twitter dataset from the #Iran network, the approach detects structural and attribute changes quickly with low false alarms. The framework provides a principled, attribute-aware method for monitoring evolving networks and is accompanied by data and code for reproducibility.

Abstract

Detecting anomalous behavior in dynamic networks remains a constant challenge. This problem is further exacerbated when the underlying topology of these networks is affected by individual highly-dimensional node attributes. We address this issue by tracking a network's modularity as a proxy of its community structure. We leverage Graph Neural Networks (GNNs) to estimate each snapshot's modularity. GNNs can account for both network structure and high-dimensional node attributes, providing a comprehensive approach for estimating network statistics. Our method is validated through simulations that demonstrate its ability to detect changes in highly-attributed networks by analyzing shifts in modularity. Moreover, we find our method is able to detect a real-world event within the \#Iran Twitter reply network, where each node has high-dimensional textual attributes.
Paper Structure (23 sections, 9 equations, 6 figures, 2 tables)

This paper contains 23 sections, 9 equations, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Dynamic Networks
  4. Deep Modularity Networks
  5. Methodology
  6. Experiments
  7. Data Generation
  8. Large Changes
  9. Structural Change
  10. Twitter Dataset
  11. Conclusion
  12. Acknowledgements
  13. Related Work
  14. Network Monitoring
  15. Temporal Community Detection
  16. ...and 8 more sections

Figures (6)

  • Figure 1: Results of synthetic experiments over varying community propensity, $\lambda_{r,r}$ and $\lambda_{r,s}$, values. Here each point represents $N=100$ simulations at their respective values of $\lambda_{r,r}$ and $\lambda_{r,s}$
  • Figure 2: Control Chart for # Iran Twitter Network when $\alpha =0.2$.
  • Figure 3: Comparison of square root collapse regularizer and collapse regularizer
  • Figure 4: Illustration of model learning during Phase I $\&$ monitoring during Phase II
  • Figure 5: Heatmap representation of varying $\lambda_{r,r}$ and $\lambda_{r,s}$ values
  • ...and 1 more figures