[Paper Review] Constant State of Change: Engagement Inequality in Temporal Dynamic Networks

Hadar Miller, Osnat Mokryn|arXiv (Cornell University)|Oct 3, 2019

Complex Network Analysis TechniquesPhysics and Astronomy68 references18 citations

TL;DR

This paper proposes a size-agnostic, online change point detection framework for evolving temporal networks that detects structural shifts by analyzing changes in degree distribution using Kolmogorov-Smirnov (KS) distance. It requires no prior knowledge of network size, node identities, or historical data, and achieves high precision (0.9) and recall (0.9) on the Enron email dataset, outperforming prior methods in both synthetic and real-world networks.

ABSTRACT

The temporal changes in complex systems of interactions have excited the research community in recent years as they encompass understandings on their dynamics and evolution. From the collective dynamics of organizations and online communities to the spreading of information and fake news, to name a few, temporal dynamics are fundamental in the understanding of complex systems. In this work, we quantify the level of engagement in dynamic complex systems of interactions, modeled as networks. We focus on interaction networks for which the dynamics of the interactions are coupled with that of the topology, such as online messaging, forums, and emails. We define two indices to capture the temporal level of engagement: the Temporal Network (edge) Intensity index, and the Temporal Dominance Inequality index. Our surprising results are that these measures are stationary for most measured networks, regardless of vast fluctuations in the size of the networks in time. Moreover, more than 80% of weekly changes in the indices values are bounded by less than 10%. The indices are stable between the temporal evolution of a network but are different between networks, and a classifier can determine the network the temporal indices belong to with high success. We find an exception in the Enron management email exchange during the year before its disintegration, in which both indices show high volatility throughout the inspected period.

Motivation & Objective

• To detect structural changes in evolving temporal networks without prior knowledge of network size or node identities.
• To develop a real-time, online framework for change point detection that is robust to varying network sizes and dynamic node sets.
• To evaluate the framework’s performance on both synthetic networks and real-world datasets like Enron and AskUbuntu.
• To compare the proposed method with existing distance metrics (KL, RH) in detecting changes in degree distributions.
• To enable sensitivity tuning for detecting gradual changes, not just abrupt shifts, through p-value-based hypothesis testing.

Proposed method

• Uses a sliding-window approach to analyze degree distributions of network snapshots over time.
• Applies the Kolmogorov-Smirnov (KS) statistical test to compare degree distributions between consecutive windows and compute a p-value indicating significance of change.
• Employs hypothesis testing with a data-driven threshold to distinguish true change points from random fluctuations.
• Does not require fitting theoretical distributions or knowing the number of nodes in advance, enabling online operation.
• Supports multiple distance metrics (KS, KL, RH) for comparison, with KS showing best performance on moderate-sized networks.
• Treats each network snapshot as a degree distribution, enabling size-agnostic analysis regardless of node count variation.

Experimental results

Research questions

• RQ1 Can a change point detection framework operate effectively without prior knowledge of network size or node identities?
• RQ2 How does the framework perform in detecting changes in degree distributions across different network models (e.g., Erdős–Rényi, small-world)?
• RQ3 What is the relative performance of KS, KL, and RH distance metrics in detecting structural changes in evolving networks?
• RQ4 Can the framework detect gradual changes over time, not just abrupt shifts, by adjusting sensitivity via p-value thresholds?
• RQ5 How does the framework compare to existing methods in precision and recall on real-world datasets with ground truth?

Key findings

• The framework achieved 90% precision and 90% recall in detecting 13 out of 14 known change points in the Enron email dataset.
• It outperformed the GHRG-based method from [3] in both precision and recall, despite the latter having no false positives.
• On synthetic Erdős–Rényi (ER) and Caveman networks, the KS-based framework showed superior detection performance compared to KL and RH metrics.
• The KS distance metric was most effective for moderate-sized networks, while RH performed better on long-tailed distributions.
• The method successfully detected changes in both abrupt and gradual trends, enabling sensitivity tuning through p-value thresholds.
• The framework is online-capable and size-agnostic, requiring no historical data or node identity tracking, making it suitable for real-time applications.