Identifying Influential and Vulnerable Nodes in Interaction Networks through Estimation of Transfer Entropy Between Univariate and Multivariate Time Series

TL;DR

This paper introduces a novel method to identify influential and vulnerable nodes in interaction networks by estimating transfer entropy between univariate and multivariate time series, improving accuracy with a new estimation scheme, and demonstrating its effectiveness on synthetic and real microbiota data.

Contribution

The paper proposes a new estimation scheme for transfer entropy that considers only significantly interacting partners, enhancing the identification of influential and vulnerable nodes in complex networks.

Findings

Successfully identifies key bacterial species in microbiota data

Demonstrates improved accuracy over naive transfer entropy estimation

Validates method with synthetic and real-world data

Abstract

Transfer entropy (TE) is a powerful tool for measuring causal relationships within interaction networks. Traditionally, TE and its conditional variants are applied pairwise between dynamic variables to infer these causal relationships. However, identifying the most influential or vulnerable node in a system requires measuring the causal influence of each component on the entire system and vice versa. In this paper, I propose using outgoing and incoming transfer entropy-where outgoing TE quantifies the influence of a node on the rest of the system, and incoming TE measures the influence of the rest of the system on the node. The node with the highest outgoing TE is identified as the most influential, or "hub", while the node with the highest incoming TE is the most vulnerable, or "anti-hub". Since these measures involve transfer entropy between univariate and multivariate time series, naive estimation methods can result in significant errors, particularly when the number of variables is comparable to or exceeds the number of samples. To address this, I introduce a novel estimation scheme that computes outgoing and incoming TE only between significantly interacting partners. The feasibility of this approach is demonstrated by using synthetic data, and by applying it to real data of oral microbiota. The method successfully identifies the bacterial species known to be key players in the bacterial community, demonstrating the power of the new method.