Introduction

  • More often that liked, one of the biggest barriers for people to use R is getting the data into it.
  • For this tutorial, we classify graph data as follows:
    • Raw R network data: Datasets with edgelist, attributes, survey data, etc.
    • Already R data: already read into R using igraph, statnet, etc.
    • Graph files: DL, UCINET, pajek, etc.
  • The netdiffuseR package includes several options to read such data.
  • This tutorial shows how to use those options (functions) so that the user generates diffnet objects (the core of netdiffuseR).
  • Importantly, while very useful, diffnet objects is not the only way to use netdiffuseR. Most of the functions can also be used with matrices and arrays.
## 
## Attaching package: 'netdiffuseR'
## The following object is masked from 'package:base':
## 
##     %*%

Raw network data

  • We call raw network data to datasets that have a somewhat raw form, for example, edgelists, adjacency matrices, survey nomination data, etc. and need to be read into R.
  • Usually this datasets are acompained with vertex attribute data.
  • The issue is how to read it into R and handle it altogether.

  • Before start, we recommend the user to take a look at the Data input functions included in the utils package (see ?read.table), to the functions included in the foreign package (useful to read from Stata, SPSS, etc.), and to the read_excel function in the readxl package1 for reading excel files into R.

Edge lists

  • In this section we review the function edgelist_to_adjmat
  • In general terms, an edgelist is a two column matrix or data frame in which each row represents a tie/link from one node to another.
  • While the basic structure of edgelists only requires having two columns (usually named ego and alter), edgelists can have more information such as intensity/weight/value of the tie, or a spell/timestamp indicating the time range during which the tie exists.
  • The function edgelist_to_adjmat supports both weights and spells.

Example 1: Loading a static diffusion network via as_diffnet

For this example we will use the fakesurvey and fakeEdgelist datasets. The later was been generated using the fakesurvey dataset, which holds survey information retrieved from 10 different individuals in two different groups. Ties in the fakeEdgelist dataset are valued, and its value coincides with the number of nominatios that each individual in the survey did to each other.

# Loading the datasets
data("fakesurvey")
data("fakeEdgelist")

Taking a look at fakesurvey’s group and id column and fakeEdgelist’s ego and alter columns the user can tell that the laters have been generated by adding up group*100 with id.

head(fakesurvey[,c("id", "group")])
#   id group
# 1  1     1
# 2  2     1
# 3  3     1
# 4  4     1
# 5  5     1
# 6  1     2
head(fakeEdgelist)
#   ego alter value
# 1 102   101     1
# 2 103   102     1
# 3 102   103     1
# 4 105   103     1
# 5 105   104     2
# 6 104   105     1

We will use this information later on to verify the way the data is sorted in the resulting diffnet objects.

To use the as_diffnet function we need at least two objects: a dynamic graph represented as either an array or a list of adjacency matrices, each of size \(n\times n\), which in our case will be \(10\times 10\), and an integer vector of size \(n=10\) which holds each vertex’s time of adoption. Lets start by generating the dynamic graph using th edgelist_to_adjmat function:

# Warning in edgelist_to_adjmat.matrix(as.matrix(edgelist), w, t0, t1, t, : Some edges a had NA/NULL value on either -times- or -w-:
#   11
# These won't be included in the adjacency matrix. The complete list will be stored as an attribute of the resulting adjacency matrix, namely, -incomplete-.

As the function warns, there is an edge that had incomplete information, and further was not used to create the adjacency matrix, the edge 11. If we take a look at that edge, we will see that indeed it had incomplete information on the weight attribute:

#    ego alter value
# 11 202  <NA>    NA

In order to address this, if we want to keep the vertex 202, an isolated vertex in the data, we need to fill that value up so that when creating the diffnet object we won’t have any problem having more attributes or times of adoption that vertices in the graph.

#    ego alter value
# 11 202  <NA>     1

As expected, there is no warning. Furthermore, we have told the function that in case of having isolated vertices to keep them, as is in the case of the edge #11 which has the vertex 202. Since we asked the function to create 5 copies of the adjacency matrix, we have a list of length 5 with adjacency matrices. Lets take a look at the first element of this list:

# 9 x 9 sparse Matrix of class "dgCMatrix"
#     101 102 103 104 105 201 202 205 210
# 101   .   .   .   .   .   .   .   .   .
# 102   1   .   1   .   .   .   .   .   .
# 103   .   1   .   .   .   .   .   .   .
# 104   .   .   .   .   1   .   .   .   .
# 105   .   .   1   2   .   .   .   .   .
# 201   .   .   .   .   .   .   .   .   .
# 202   .   .   .   .   .   .   .   .   .
# 205   .   .   .   .   .   1   .   .   1
# 210   .   .   .   .   .   1   .   1   .

As you can see, the edgelist_to_adjmat function kept the vertices labels and included them as dimnames in the matrix.2 Now that our adjacency matrix has the number of elements that we expected, which actually coincides with the number of rows in the fakesurvey dataset, we can create a diffnet object:

# Dynamic network of class -diffnet-
#  Name               : Diffusion Network
#  Behavior           : Unspecified
#  # of nodes         : 9 (101, 102, 103, 104, 105, 201, 202, 205, ...)
#  # of time periods  : 5 (1 - 5)
#  Type               : directed
#  Final prevalence   : 0.89
#  Static attributes  : id, toa, group, net1, net2, net3, age, gender, not... (9)
#  Dynamic attributes : -

Example 2: Loading a static diffusion network via edgelist_to_diffnet

Following the previous example, instead of “manually” generating the adjacency matrix and calling the as_diffnet function, we will use the edgelist_to_diffnet function. The most important issue when calling this routine is to have matching ids between the edgelist and the attributes dataset. So before calling the edgelist_to_diffnet function we need to fix the id column in the fakesurvey dataset:3

# [1]  1  2  3  4  5  1  2  5 10
# [1] 101 102 103 104 105 201 202 205 210

Now that it is fixed, we can call the edgelist_to_diffnet function

# Warning in check_var_class_and_coerce(x, edgelist, c("factor", "integer", :
# Coercing -ego- into character.
# Warning in check_var_class_and_coerce(x, edgelist, c("factor", "integer", :
# Coercing -alter- into character.
# Dynamic network of class -diffnet-
#  Name               : Diffusion Network
#  Behavior           : Unspecified
#  # of nodes         : 9 (101, 102, 103, 104, 105, 201, 202, 205, ...)
#  # of time periods  : 5 (1 - 5)
#  Type               : directed
#  Final prevalence   : 0.89
#  Static attributes  : group, net1, net2, net3, age, gender, note (7)
#  Dynamic attributes : -

As a difference with the previous example, here the algorithm makes sure that the ordering of the dataset and the vertices in the adjacency matrix coincide. The previous example did gave us a correctly sorted diffnet object, but that may not always be the case. Nevertheless, the option id.and.per.vars allows the user providing with the names of the variables in the vertex attribute datasets that hold the ids and time period ids of each observation, so that the function sorts the data before coercing it into diffnet objects. More on this in the following examples.

Survey data: Nomination networks

  • In this part we review the function survey_to_diffnet
  • This function can use as input either a longitudinal dataset (which should be in long format, this is, one row per individual and time period), or a cross sectional dataset.
  • For this example we will use both fakesurvey, which holds cross section data, and fakesurveyDyn, which holds longitudinal data.

Example 3: Cross section diffusion data

We start by taking a look at the data

#   id toa group net1 net2 net3 age gender                                   note
# 1  1   1     1   NA   NA   NA  30      M                         No nominations
# 2  2   5     1    3    1   NA  35      F                          Nothing weird
# 3  3   5     1   NA    2   NA  31      F                 Only nominates in net2
# 4  4   3     1    6    5   NA  30      M Nominates someone who wasn't interview
# 5  5   2     1    4    4    3  40      F                  Nominates 4 two times
# 6  1   4     2    3    4    8  29      F               Only nominates outsiders
# 7  2   3     2    3   NA   NA  35      M                               Isolated
# 8  5   3     2   10    1   NA  50      M                          Nothing weird
# 9 10  NA     2    5    1   NA  19      F                            Non-adopter

A couple of important remarks for this dataset. First, each individual in this dataset belongs to a different group, while this is not always the case, survey_to_diffnet allows accounting for this through the groupvar argument. Also, besides of having an isolated vertex, two individuals in the survey nominate people that neither weren’t survey nor show in their groups:

#   id toa group net1 net2 net3 age gender                                   note
# 4  4   3     1    6    5   NA  30      M Nominates someone who wasn't interview
# 6  1   4     2    3    4    8  29      F               Only nominates outsiders

So in group one 4 nominates id 6, who does not show in the data, and in group two 6 nominates 3, 4, and 8, also individuals who don’t show up in the survey.

While for some researchers nominations of unsurveyed individuals may not be of importance, for some others might be. For such cases, the function has the option of either keeping unsurveyed individuals (so you would get a bigger adjacency matrix), or ignore them and keep only those who were surveyed. For example, if we wanted to keep unsurveyed individuals in the network we would need to set no.unsurveyed = FALSE:

# Dynamic network of class -diffnet-
#  Name               : Diffusion Network
#  Behavior           : Unspecified
#  # of nodes         : 13 (101, 102, 103, 104, 105, 106, 201, 202, ...)
#  # of time periods  : 5 (1 - 5)
#  Type               : directed
#  Final prevalence   : 0.62
#  Static attributes  : group, net1, net2, net3, age, gender, note (7)
#  Dynamic attributes : -
# Retrieving nodes ids
nodes(diffnet_w_unsurveyed)
#  [1] "101" "102" "103" "104" "105" "106" "201" "202" "203" "204" "205" "208"
# [13] "210"

A network spanning 5 time periods with 13 vertices (9 surveyed individuals + 4 unsurveyed individuals). This produces a different result when compared to the case in which me use the default behavior of the function, no.unsurveyed = TRUE:

# Dynamic network of class -diffnet-
#  Name               : Diffusion Network
#  Behavior           : Unspecified
#  # of nodes         : 9 (101, 102, 103, 104, 105, 201, 202, 205, ...)
#  # of time periods  : 5 (1 - 5)
#  Type               : directed
#  Final prevalence   : 0.89
#  Static attributes  : group, net1, net2, net3, age, gender, note (7)
#  Dynamic attributes : -
# Retrieving nodes ids
nodes(diffnet_wo_unsurveyed)
# [1] "101" "102" "103" "104" "105" "201" "202" "205" "210"

Furthermore, we can compare the two diffusion networks by sustracting one from another:

# Dynamic network of class -diffnet-
#  Name               : Diffusion Network
#  Behavior           : Unspecified
#  # of nodes         : 4 (106, 203, 204, 208)
#  # of time periods  : 5 (1 - 5)
#  Type               : directed
#  Final prevalence   : 0.00
#  Static attributes  : group, net1, net2, net3, age, gender, note (7)
#  Dynamic attributes : -

Example 4: Longitudinal diffusion data

In this example we will use dynamic network data, this is, an edgelist with spells and dynamic attributes

# Taking a look at the data
data("fakeDynEdgelist")
head(fakeDynEdgelist)
#   ego alter value time
# 1 102   101     1 1990
# 2 103   102     1 1990
# 3 102   103     1 1990
# 4 105   103     1 1990
# 5 105   104     2 1990
# 6 104   105     1 1990
data("fakesurveyDyn")
head(fakesurveyDyn)
#   id  toa group net1 net2 net3 age gender
# 1  1 1991     1   NA   NA   NA  30      M
# 2  2 1990     1    3    1   NA  35      F
# 3  3 1991     1   NA    2   NA  31      F
# 4  4 1990     1    6    5   NA  30      M
# 5  5 1991     1    4    4    3  40      F
# 6  1 1991     2    3    4    8  29      F
#                                                 note time
# 1                         First wave: No nominations 1990
# 2                          First wave: Nothing weird 1990
# 3                 First wave: Only nominates in net2 1990
# 4 First wave: Nominates someone who wasn't interview 1990
# 5                  First wave: Nominates 4 two times 1990
# 6               First wave: Only nominates outsiders 1990

Same as before, we have to make sure the ids are right

# Warning in check_var_class_and_coerce(x, dat, c("numeric", "integer"),
# "integer", : Coercing -time- into integer.
# Warning in check_var_class_and_coerce(x, edgelist, c("factor", "integer", :
# Coercing -ego- into character.
# Warning in check_var_class_and_coerce(x, edgelist, c("factor", "integer", :
# Coercing -alter- into character.
# Dynamic network of class -diffnet-
#  Name               : Diffusion Network
#  Behavior           : Unspecified
#  # of nodes         : 9 (101, 102, 103, 104, 105, 201, 202, 205, ...)
#  # of time periods  : 2 (1990 - 1991)
#  Type               : directed
#  Final prevalence   : 1.00
#  Static attributes  : -
#  Dynamic attributes : group, net1, net2, net3, age, gender, note (7)

  1. While there are other candidates as the openxlsx package, the readxl package has the nice feature of correctly processing the encoding of the excel files. This is specially important if you are dealing with non ASCII or UTF-8 datasets.

  2. Another thing to tell, the matrices stored in adjmat are of class dgCMatrix from the Matrix package, these are Column Compressed Stored sparse matrices and allows saving memory in matrices with many zeros. netdiffuseR routines are based in this class of matrices. Furthermore, to have an idea of how much memory sparse matrices save, while a square matrix of size \(5e4\times 5e4\) would need close to 18GB of memory using a regular R matrix, a dgCMatrix of the same size takes around 6MB.

  3. The with function allows simplifying data management in R by allowing to reference columns in a data.frame without having to call the data.frame itself (see ?with).