Measuring the Hierarchy of Feedforward Networks

TL;DR

This paper introduces a quantitative hierarchical index for feedforward networks based on graph and information theory, enabling classification of structures as hierarchical, anti-hierarchical, or non-hierarchical.

Contribution

It defines a novel hierarchical index using entropy measures to quantify hierarchy in directed acyclic graphs, distinguishing different causal structures.

Findings

Maximal hierarchy in feedforward trees and inverted trees identified.

Null hierarchy values correspond to linear chains and fully connected graphs.

The index effectively classifies various causal network structures.

Abstract

In this paper we explore the concept of hierarchy as a quantifiable descriptor of ordered structures, departing from the definition of three conditions to be satisfied for a hierarchical structure: {\em order}, {\em predictability} and {\em pyramidal structure}. According to these principles we define a hierarchical index taking concepts from graph and information theory. This estimator allows to quantify the hierarchical character of any system susceptible to be abstracted in a feedforward causal graph, i.e., a directed acyclic graph defined in a single connected structure. Our hierarchical index is a balance between this predictability and pyramidal condition by the definition of two entropies: one attending the onward flow and other for the backward reversion. We show how this index allows to identify hierarchical, anti-hierarchical and non hierarchical structures. Our formalism reveals that departing from the defined conditions for a hierarchical structure, feedforward trees and the inverted tree graphs emerge as the only causal structures of maximal hierarchical and anti-hierarchical systems, respectively. Conversely, null values of the hierarchical index are attributed to a number of different configuration networks; from linear chains, due to their lack of pyramid structure, to full-connected feedforward graphs where the diversity of onward pathways is canceled by the uncertainty (lack of predictability) when going backwards. Some illustrative examples are provided for the distinction among these three types of hierarchical causal graphs.