Coevolution of Network Structure and Content

As individuals communicate, their exchanges form a dynamic network. We demonstrate, using time series analysis of communication in three online settings, that network structure alone can be highly revealing of the diversity and novelty of the information being communicated. Our approach uses both standard and novel network metrics to characterize how unexpected a network configuration is, and to capture a network’s ability to conduct information. We find that networks with a higher conductance in link structure exhibit higher information entropy, while unexpected network configurations can be tied to information novelty. We use a simulation model to explain the observed correspondence between the evolution of a network’s structure and the information it carries.

💡 Research Summary

The paper investigates the co‑evolution of network structure and the content that flows through it, using time‑series analysis of three distinct online communication environments: Twitter, Second Life (SL), and the Enron email corpus. The authors argue that the dynamic topology of a communication network alone can reveal key properties of the information being exchanged, specifically its diversity (how many different topics are present) and its novelty (how much the current topics differ from those in the recent past).

Data preparation involves segmenting each communication stream into fixed‑size windows measured by the number of actions rather than clock time (800 tweets per segment for Twitter, 50 asset transfers per segment for SL, and 100 emails per segment for Enron). Within each segment a directed graph is constructed where nodes are participants and edges represent at least one interaction during that window.

For each segment the authors compute a suite of standard network metrics (node count, edge count, reciprocity, clustering coefficient, degree statistics, size of the largest strongly connected component, degree assortativity, and a Gini‑based centralization measure). In addition they introduce two novel descriptors: (1) Cycle‑free effective conductance, which aggregates over all node pairs the sum of inverse‑degree‑weighted path contributions, thereby quantifying the capacity of the network to transmit information; and (2) Expectedness, defined as the average of the previous‑segment conductance values for the edges that appear in the current segment, capturing how “surprising” the current edge set is given recent topology.

Content characteristics are measured differently for each platform. In SL the diversity of assets exchanged is captured by the Shannon entropy of the asset‑type distribution, while novelty is measured by the Jaccard similarity of asset sets across consecutive windows. For textual data (Twitter and Enron) each message is transformed into a TF‑IDF weighted word‑frequency vector; cosine similarity between vectors from consecutive windows provides a measure of content change, and the entropy of the word‑frequency distribution within a window serves as a proxy for diversity.

Correlation analysis across the time series reveals consistent patterns: higher conductance correlates positively with information entropy, indicating that more “conductive” networks tend to host a broader set of topics. Conversely, lower expectedness (i.e., more unexpected edge configurations) correlates strongly with higher novelty scores, suggesting that sudden structural shifts accompany the emergence of new topics. These relationships hold across all three datasets, though the magnitude of correlations varies (e.g., conductance‑entropy correlation ≈0.58 in Twitter, weaker but still significant in SL and Enron). Traditional metrics such as clustering or reciprocity are inter‑correlated but do not explain content dynamics as cleanly as conductance and expectedness.

To explain the observed co‑evolution, the authors present an agent‑based simulation. Agents possess pieces of information with an intrinsic “value”. High‑value information prompts agents to form new connections, thereby increasing network conductance; low‑value information is routed through existing paths. The simulation reproduces the empirical findings: periods of high conductance coincide with high entropy, and abrupt changes in edge sets (low expectedness) coincide with spikes in novelty.

The paper’s contributions are threefold: (1) Demonstrating that dynamic network topology alone can predict statistical properties of the underlying information; (2) Introducing conductance and expectedness as novel, interpretable metrics linking structure to flow; (3) Providing a generalizable analytical framework applicable to heterogeneous online platforms. The findings have practical implications for real‑time trend detection, crisis monitoring, and the design of information‑diffusion interventions, as they suggest that monitoring structural “surprises” can serve as early warnings of emerging topics or events.