Positional dominance in networks

This vignette describes the concept of positional dominance, the generalization of neighborhood-inclusion for arbitrary network and attribute data. Additionally, some use cases with the netrankr package are given. The partial ranking induced by positional dominance can be used to assess partial centrality or compute probabilistic centrality.

Theoretical Background

A network can be described as a dyadic variable \(x\in \mathcal{W}^\mathcal{D}\), where \(\mathcal{W}\) is the value range of the network (in the simple case of unweighted networks \(\mathcal{W}=\{0,1\}\)) and \(\mathcal{D}=\mathcal{N}\times\mathcal{A}\) describes the dyadic domain of actors \(\mathcal{N}\) and affiliations \(\mathcal{A}\). If \(\mathcal{A}\neq\mathcal{N}\), we obtain a two-mode network and if \(\mathcal{A}=\mathcal{N}\) a one-mode network consisting of relations among actors.


Let \(x\in \mathcal{W}^\mathcal{D}\) be a network and \(i,j \in \mathcal{N}\). We say that \(i\) is dominated by \(j\) under the total homogeneity assumption, denoted by \(i \leq j\) if \[ x_{it}\leq x_{jt} \quad \forall t \in \mathcal{N}. \] If there exists a permutation \(\pi: \mathcal{N} \to \mathcal{N}\) such that \[ x_{it}\leq x_{j\pi(t)} \quad \forall t \in \mathcal{N}, \] we say that \(i\) is dominated by \(j\) under the total heterogeneity assumption, denoted by \(i ⪯ j\).
It holds that \(i\leq j \implies i ⪯ j\) but not vice versa.

More about the positional dominance and the positional approach to network analysis can be found in

Brandes, Ulrik. (2016). Network Positions. Methodological Innovations, 9, 2059799116630650. (link)

Positional Dominance in the netrankr Package

set.seed(1886) #for reproducibility

The function positional_dominance can be used to check both, dominance under homogeneity and heterogeneity. In accordance with the analytic pipeline of centrality we use the %>% operator.

g <- dbces11

#neighborhood inclusion can be expressed with the analytic pipeline
D <- g %>% indirect_relations(type="adjacency") %>% positional_dominance()

More on the indirect_relations() function can be found in this vignette.

The map parameter of positional_dominance allows to distinguish between dominance under total heterogeneity (map=FALSE) and total homogeneity (map=TRUE). In the later case, all relations can be ordered non-decreasingly (or non-increasingly if the relation describes costs, such as distances) and afterwards checked front to back. Dominance under total homogeneity yields a ranking, if the relation is binary (e.g. adjacent or not).

D <- g %>% 
  indirect_relations(type="adjacency") %>% 

## [1] 1

For cost variables like shortest path distances, the benefit parameter is set to FALSE.

D1 <- g %>% 
  indirect_relations(type="dist_sp") %>% 

From the definition given in the first section, it is clear that there are always at least as many comparable pairs under the total homogeneity assumption as under total heterogeneity.

D1 <- g %>% 
  indirect_relations(type="dist_sp") %>% 

D2 <- g %>% 
  indirect_relations(type="dist_sp") %>% 

## heterogeneity   homogeneity 
##     0.1636364     0.8727273

Additionally, all dominance relations from the heterogeneity assumption are preserved under total homogeneity. (Note: \(A\implies B\) is equivalent to \(\neg A \lor B\))

all(D1!=1 | D2==1) 
## [1] TRUE