[Paper Review] Dynamic Network Model from Partial Observations

Elahe Ghalebi, Baharan Mirzasoleiman+2 more|arXiv (Cornell University)|May 27, 2018

Complex Network Analysis Techniques26 references9 citations

TL;DR

This paper proposes Dyference, a non-parametric generative dynamic network model that infers time-varying network structures from partial observations of cascade diffusion—specifically, node infection times—using a mixture of coupled hierarchical Dirichlet processes. The method enables online inference of evolving community structures and predictive edge probabilities, including unobserved or future edges, and outperforms existing methods in dynamic network inference, diffusion prediction, and bankruptcy prediction on real-world and synthetic networks.

ABSTRACT

Can evolving networks be inferred and modeled without directly observing their nodes and edges? In many applications, the edges of a dynamic network might not be observed, but one can observe the dynamics of stochastic cascading processes (e.g., information diffusion, virus propagation) occurring over the unobserved network. While there have been efforts to infer networks based on such data, providing a generative probabilistic model that is able to identify the underlying time-varying network remains an open question. Here we consider the problem of inferring generative dynamic network models based on network cascade diffusion data. We propose a novel framework for providing a non-parametric dynamic network model--based on a mixture of coupled hierarchical Dirichlet processes-- based on data capturing cascade node infection times. Our approach allows us to infer the evolving community structure in networks and to obtain an explicit predictive distribution over the edges of the underlying network--including those that were not involved in transmission of any cascade, or are likely to appear in the future. We show the effectiveness of our approach using extensive experiments on synthetic as well as real-world networks.

Motivation & Objective

• To address the open problem of generating generative probabilistic models for time-varying networks when only partial cascade diffusion data (e.g., infection times) is available.
• To develop a framework that infers not only likely edges but also unobserved or future edges, and evolving community structures.
• To provide online, time-varying estimates of edge probabilities without requiring prior knowledge of network structure or transmission rates.
• To enable accurate prediction of diffusion cascades and influential nodes in dynamic networks using only infection time data.

Proposed method

• Leverages a mixture of coupled hierarchical Dirichlet processes (MDND) to model edge-exchangeable, non-parametric dynamic networks with evolving community structure.
• Processes cascade diffusion data in time intervals, computing marginal edge probabilities from inferred diffusion trees for each cascade.
• Uses a Gibbs sampler to iteratively update model variables based on sampled edges and updated edge probability distributions.
• Employs an iterative inference loop: update edge probabilities from cascade data, re-estimate diffusion tree distributions, and refine model parameters until convergence.
• Extends to arbitrary generative network models beyond MDND, enabling flexibility in latent representation choices.
• Supports online inference by processing data in temporal windows, enabling real-time tracking of network evolution.

Experimental results

Research questions

• RQ1 Can a generative probabilistic model for dynamic networks be constructed from only partial observations of cascade diffusion, such as node infection times?
• RQ2 Can such a model infer not only likely edges but also unobserved or future edges and evolving community structures?
• RQ3 How accurately can the model predict future diffusion cascades and influential nodes when no network structure is known?
• RQ4 How does the model compare to existing methods in terms of prediction accuracy and computational efficiency on real-world and synthetic networks?

Key findings

• Dyference outperforms InfoPath and NetRate in online dynamic network inference on time-varying hyperlink and social media networks, achieving higher F1 scores across all evaluated datasets.
• On the financial transaction network, Dyference significantly outperforms InfoPath in predicting future bankruptcies, with improved Map@k and Hits@k metrics over 2-year windows.
• Dyference achieves competitive running time compared to InfoPath while consistently surpassing it in prediction accuracy, demonstrating scalability and efficiency.
• The model successfully predicts influential websites on LinkedIn using only infection times, with results comparable to those from known network structures.
• In diffusion prediction tasks, Dyference achieves strong performance on Map@k and Hits@k without requiring prior knowledge of network structure or transmission rates.
• The framework enables accurate tracking of evolving community structures and provides explicit predictive distributions over all potential edges, including those not involved in any observed cascade.