Switch branches/tags
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
214 lines (173 sloc) 8.28 KB
title: "netrankr: Centrality (almost) without Indices"
author: ""
output: github_document
always_allow_html: yes
[![CRAN Status Badge](](
[![CRAN Downloads](](
[![Travis-CI Build Status](](
```{r setup, include=FALSE}
echo = TRUE,
comment = "#>",
fig.path = "index-",
fig.width = 6,
fig.align = 'center',
out.width = "70%")
# Overview
The literature is flooded with centrality indices and new ones are introduced
on a regular basis. Although there exist several theoretical and empirical guidelines
on when to use certain indices, there still exists plenty of ambiguity in the concept
of network centrality. To date, network centrality is nothing more than applying indices
to a network:
```{r old,echo=FALSE}
DiagrammeR::DiagrammeR('graph LR
A[network]==>B{centrality <br> indices}
style A fill:#fff,stroke:#000, stroke-width:2px
style B fill:#fff,stroke:#000, stroke-width:2px
', height = 200
The only degree of freedom is the choice of index. The package comes with an Rstudio addin (`index_builder()`),
which allows to build or choose from more than 20 different indices. Blindly (ab)using
this function is highly discouraged!
The `netrankr` package is based on the idea that centrality is more than a
conglomeration of indices. Decomposing them in a series of microsteps offers
the posibility to gradually add ideas about centrality, without succumbing to
trial-and-error approaches. Further, it allows for alternative assessment methods
which can be more general than the index-driven approach:
```{r new,echo=FALSE}
DiagrammeR::DiagrammeR('graph LR
A[network]==>B{indirect relation}
B ==>P[position]
P==>C[aggregate position]
C==>D[centrality index]
P==>E{positional <br> dominance}
E==>F[partial centrality]
E==>G[probabilistic centrality]
style A fill:#fff,stroke:#000, stroke-width:2px
style B fill:#fff,stroke:#000, stroke-width:2px
style C fill:#fff,stroke:#000, stroke-width:2px
style D fill:#fff,stroke:#000, stroke-width:2px
style E fill:#fff,stroke:#000, stroke-width:2px
style F fill:#fff,stroke:#000, stroke-width:2px
style G fill:#fff,stroke:#000, stroke-width:2px
style P fill:#EEC900,stroke:#000, stroke-width:3px
', height = 300
The new approach is centered around the concept of *positions*, which are defined as
the relations and potential attributes of a node in a network. The aggregation
of the relations leads to the definition of indices. However, positions can also
be compared via *positional dominance*, leading to partial centrality rankings and
the option to calculate probabilistic centrality rankings.
For a more detailed theoretical background, consult the [Literature](#literature)
at the end of this page.
## Installation
To install from CRAN:
```{r install_cran, eval=FALSE}
To install the developer version from github:
```{r install_git, eval=FALSE}
## Simple Example
This example briefly explains some of the functionality of the package and the
difference to an index driven approach. For a more realistic application see
the [use case](articles/use_case.html) example.
We work with the following small graph.
```{r example_graph, warning=FALSE,message=FALSE}
g <- graph.empty(n = 11,directed = FALSE)
g <- add_edges(g,c(1,11,2,4,3,5,3,11,4,8,5,9,5,11,6,7,6,8,
```{r dbces_neutral, echo=FALSE}
Say we are interested in the most central node of the graph and simply compute some
standard centrality scores with the `igraph` package. Defining centrality indices
in the `netrankr` package is explained [here](articles/centrality_indices.html).
```{r cent,warning=FALSE}
cent_scores <- data.frame(
degree = degree(g),
betweenness = round(betweenness(g),4),
closeness = round(closeness(g),4),
eigenvector = round(eigen_centrality(g)$vector,4),
subgraph = round(subgraph_centrality(g),4))
# What are the most central nodes for each index?
```{r dbces_color, echo=FALSE}
As you can see, each index assigns the highest value to a different vertex.
A more general assessment starts by calculating the neighborhood inclusion preorder.
```{r ex_ni}
P <- neighborhood_inclusion(g)
[Schoch & Brandes (2016)]( showed that
$N(u) \subseteq N[v]$ (i.e. `P[u,v]=1`) implies $c(u) \leq c(v)$ for
centrality indices $c$, which are defined via specific path algebras. These include
many of the well-known measures like closeness (and variants), betweenness (and variants)
as well as many walk-based indices (eigenvector and subgraph centrality, total communicability,...).
Neighborhood-inclusion defines a partial ranking on the set of nodes. Each ranking
that is in accordance with this partial ranking yields a proper centrality ranking.
Each of these ranking can thus potentially be the outcome of a centrality index.
Using rank intervals, we can examine the minimal and maximal possible rank of each node.
The bigger the intervals are, the more freedom exists for indices to rank nodes differently.
```{r partial}
plot_rank_intervals(P,cent.df = cent_scores,ties.method="average")
The potential ranks of nodes are not uniformly distributed in the intervals. To get
the exact probabilities, the function `exact_rank_prob()` can be used.
```{r ex_p}
res <- exact_rank_prob(P)
`lin.ext` is the number of possible rankings. For the graph `g` we could therefore come up with
`r format(res$lin.ext,big.mark = ",")` indices that would rank the nodes differently.
`rank.prob` contains the probabilities for each node to occupy a certain rank.
For instance, the probability for each node to be the most central one is as follows.
```{r most_central}
round(res$rank.prob[ ,11],2)
`relative.rank` contains the relative rank probabilities. An entry `relative.rank[u,v]`
indicates how likely it is that `v` is more central than `u`.
```{r rel_rank}
# How likely is it, that 6 is more central than 3?
`expected.ranks` contains the expected centrality ranks for all nodes. They are
derived on the basis of `rank.prob`.
```{r exp_rank}
The higher the value, the more central a node is expected to be.
**Note**: The set of rankings grows exponentially in the number of nodes and the exact
calculation becomes infeasible quite quickly and approximations need to be used.
Check the [benchmark results](articles/benchmarks.html) for guidelines.
## Theoretical Background {#literature}
`netrankr` is based on a series of papers that appeared in recent years. If you
want to learn more about the theoretical background of the package,
consult the following literature:
> Schoch, David. (2018). Centrality without Indices: Partial rankings and rank
Probabilities in networks. *Social Networks*, **54**, 50-60.([link](
> Schoch, David & Valente, Thomas W., & Brandes, Ulrik. (2017). Correlations among centrality indices
and a class of uniquely ranked graphs. *Social Networks*, **50**, 46-54.([link](
> Schoch, David & Brandes, Ulrik. (2016). Re-conceptualizing centrality in social networks.
*European Journal of Appplied Mathematics*, **27**(6), 971–985.
> Brandes, Ulrik. (2016). Network Positions.
*Methodological Innovations*, **9**, 2059799116630650.