Setting up
For this session, we’re going to use the “ison_algebra” dataset
included in the {manynet}
package. Do you remember how to
call the data? Can you find out some more information about it via its
help file?
# Let's call and load the 'ison_algebra' dataset
data("ison_algebra", package = "manynet")
# Or you can retrieve like this:
ison_algebra <- manynet::ison_algebra
# If you want to learn more about the 'ison_algebra' dataset, use the following function (below)
?manynet::ison_algebra
data("ison_algebra", package = "manynet")
?manynet::ison_algebra
# If you want to see the network object, you can run the name of the object
# ison_algebra
# or print the code with brackets at the front and end of the code
# (ison_algebra <- manynet::ison_algebra)
We can see that the dataset is multiplex, meaning that it contains several different types of ties: friendship (friends), social (social) and task interactions (tasks).
Separating multiplex networks
As a multiplex network, there are actually three different types of
ties in this network. We can extract them and investigate them
separately using to_uniplex()
. Within the parentheses, put
the multiplex object’s name, and then as a second argument put the name
of the tie attribute in quotation marks. Once you have extracted all
three networks, graph them and add a descriptive title.
# Here's the basic idea/code syntax you will need to extract each type of network
# You will want to replace
____ <- to_uniplex(ison_algebra, _____)
# Now, let's compare the each attribute's graph, side-by-side by using "+"
# Note: using "/" after each graph will order them vertically; however, it might not be best way
# See for example:
gfriend <- autographr(friends) + ggtitle("Friendship")
gfriend + gsocial + gtask
friends <- to_uniplex(ison_algebra, "friends")
gfriend <- autographr(friends) + ggtitle("Friendship")
social <- to_uniplex(ison_algebra, "social")
gsocial <- autographr(social) + ggtitle("Social")
tasks <- to_uniplex(ison_algebra, "tasks")
gtask <- autographr(tasks) + ggtitle("Task")
gfriend + gsocial + gtask
Note also that these are weighted networks. autographr()
automatically recognises these different weights and plots them.
Structural Holes and Constraint
Our first question for this network, is where innovation and creative ideas might be expected to appear.
Measuring structural holes
Let’s take a look at which actors are least constrained by
their position in the task network to begin with.
{migraph}
makes this easy enough with the
node_constraint()
function.
node_constraint(____)
# Don't forget we want to look at which actors are least constrained by their position in the 'tasks' network
node_constraint(tasks)
This function returns a vector of constraint scores that can range
between 0 and 1. Let’s graph the network again, sizing the nodes
according to this score. We can also identify the node with the minimum
constraint score using node_is_min()
.
tasks <- tasks %>%
mutate(constraint = node_constraint(____),
low_constraint = node_is_min(node_constraint(____)))
# Don't forget, we are still looking at the 'tasks' network
# Now, let's graph the network
# Note 1: we are looking at the 'tasks' network
# Note 2: we are interested in the actors 'least constrained' by their position
autographr(____, node_color = "____")
autographr(tasks, node_size = "constraint", node_color = "low_constraint")
tasks <- tasks %>%
mutate(constraint = node_constraint(tasks),
low_constraint = node_is_min(node_constraint(tasks)))
autographr(tasks, node_size = "constraint", node_color = "low_constraint")
Why minimum? Because constraint measures how well connected each node’s partners are, with the implication that having few partners that are already connected to each other puts a node in an advantageous position to identify and share novel solutions to problems. So what can we learn from this plot about where innovation might occur within this network?
Structural Equivalence
Next we might ask ourselves what (other) roles there are in the network? We want to know who plays what role in this algebra class. Let us begin with structural equivalence.
We’re going to identify structurally equivalent positions across all
the data that we have, including ‘task’, ‘social’, and ‘friend’ ties. So
that is, we are using the multiplex ison_algebra
dataset
again and not a uniplex subgraph thereof.
Finding structurally equivalent classes
In {migraph}
, finding how the nodes of a network can be
partitioned into structurally equivalent classes can be as easy as:
node_structural_equivalence(ison_algebra)
ison_algebra %>%
mutate(se = node_structural_equivalence(ison_algebra)) %>%
autographr(node_color = "se")
But actually, a lot is going on behind the scenes here that we can unpack. Understanding what is going on behind the scenes is important for understanding how these classes are identified and how to interpret them.
Step one: starting with a census
All equivalence classes are based on nodes’ similarity across some
profile of motifs. In {migraph}
, we call these motif
censuses. Any kind of census can be used, and
{migraph}
includes a few options, but
node_structural_equivalence()
is based off of the census of
all the nodes’ ties, both outgoing and incoming ties, to characterise
their relationships to tie partners.
# Let's use the node_tie_census() function
# The function accepts an object such as a dataset
# Hint: Which dataset are we using in this tutorial?
node_tie_census(____)
node_tie_census(ison_algebra)
# Now, let's get the dimensions of an object via the dim() function
dim(node_tie_census(ison_algebra))
node_tie_census(ison_algebra)
dim(node_tie_census(ison_algebra))
We can see that the result is a matrix of 16 rows and 96 columns, because we want to catalogue or take a census of all the different incoming/outgoing partners our 16 nodes might have across these three networks. Note also that the result is a weighted matrix; what would you do if you wanted it to be binary?
# we could convert the result using as.matrix, returning the ties
as.matrix((node_tie_census(ison_algebra)>0)+0)
# But it's easier to simplify the network by removing the classification into different types of ties.
# Note that this also reduces the total number of possible paths between nodes
ison_algebra %>%
select_ties(-c(friends, social, tasks)) %>%
node_tie_census()
Note that node_tie_census()
does not need to be passed
to node_structural_equivalence()
— this is done
automatically! However, the more generic node_equivalence()
is available and can be used with whichever tie census is desired. Feel
free to explore using some of the other censuses available in
{migraph}
, though some common ones are already used in the
other equivalence convenience functions,
e.g. node_triad_census()
in
node_regular_equivalence()
and
node_path_census()
in
node_automorphic_equivalence()
.
Step two: growing a tree of similarity
The next part takes this census and creates a dendrogram based on
distance or dissimilarity among the nodes’ census profiles. This is all
done internally within e.g. node_structural_equivalence()
,
though there are two important parameters that can be set to obtain
different results.
First, users can set the type of distance measure used. For
enthusiasts, this is passed on to stats::dist()
, so that
help page should be consulted for more details. By default
"euclidean"
is used.
Second, we can also set the type of clustering algorithm employed. By
default, {migraph}
’s equivalence functions use hierarchical
clustering, "hier"
, but for compatibility and enthusiasts,
we also offer "concor"
, which implements a CONCOR
(CONvergence of CORrelations) algorithm.
We can see the difference from varying the clustering algorithm
and/or distance by plotting the dendrograms (hidden) in the output from
node_structural_equivalence()
:
alge <- to_named(ison_algebra) # fake names to make comparison clearer
plot(node_structural_equivalence(alge,
cluster = "hier", distance = "euclidean"))
# changing the type of distance used
plot(node_structural_equivalence(alge,
cluster = "hier", distance = "manhattan"))
# changing the clustering algorithm
plot(node_structural_equivalence(alge,
cluster = "concor", distance = "euclidean"))
So plotting a membership
vector from
{migraph}
returns a dendrogram with the names of the nodes
on the y-axis and the distance between them on the
x-axis. Using the census as material, the distances between the
nodes is used to create a dendrogram of (dis)similarity among the nodes.
Basically, as we move to the right, we’re allowing for more and more
dissimilarity among those we cluster together. A fork or branching point
indicates the level of dissimilarity at which those two or more nodes
would be said to be equivalent. Where two nodes’ branches join/fork
represents the maximum distance among all their leaves, so more similar
nodes’ branches fork closer to the tree’s canopy, and less similar
(groups of) nodes don’t join until they form basically the trunk.
Note that with the results using the hierarchical clustering algorithm, the distance directly affects the structure of the tree (and the results).
The CONCOR dendrogram operates a bit differently to hierarchical clustering though. Instead it represents how converging correlations repeatedly bifurcate the nodes into one of two partitions. As such the ‘distance’ is really just the (inverse) number of steps of bifurcations until nodes belong to the same class.
Step three: identifying the number of clusters
Another bit of information represented in the dendrogram is where the tree should be cut (the dashed red line) and how the nodes are assigned to the branches (clusters) present at that cut-point.
But where does this red line come from? Or, more technically, how do we identify the number of clusters into which to assign nodes?
{migraph}
includes several different ways of
establishing k
, or the number of clusters. Remember, the
further to the right the red line is (the lower on the tree the cut
point is) the more dissimilar we’re allowing nodes in the same cluster
to be. We could set this ourselves by just passing k
an
integer.
plot(node_structural_equivalence(alge, k = 2))
But we’re really just guessing. Maybe 2 is not the best
k
? To establish what the best k
is for this
clustering exercise, we need to iterate through a number of potential
k
and consider their fitness by some metric. There are a
couple of options here.
One is to consider, for each k
, how correlated this
partition is with the observed network. When there is one cluster for
each vertex in the network, cell values will be identical to the
observed correlation matrix, and when there is one cluster for the whole
network, the values will all be equal to the average correlation across
the observed matrix. So the correlations in each by-cluster matrix are
correlated with the observed correlation matrix to see how well each
by-cluster matrix fits the data.
Of course, the perfect partition would then be where all nodes are in
their own cluster, which is hardly ‘clustering’ at all. Also, increasing
k
will always improve the correlation. But if one were to
plot these correlations as a line graph, then we might expect there to
be a relatively rapid increase in correlation as we move from, for
example, 3 clusters to 4 clusters, but a relatively small increase from,
for example, 13 clusters to 14 clusters. By identifying the inflection
point in this line graph, {migraph}
selects a number of
clusters that represents a trade-off between fit and parsimony. This is
the k = "elbow"
method.
The other option is to evaluate a candidate for k
based
not on correlation but on a metric of how similar each node in a cluster
is to others in its cluster and how dissimilar each node is to
those in a neighbouring cluster. When averaged over all nodes and all
clusters, this provides a ‘silhouette coefficient’ for a candidate of
k
. Choosing the number of clusters that maximizes this
coefficient, which is what k = "silhouette"
does, can
return a somewhat different result to the elbow method. See what we have
here, with all other arguments held the same:
plot(node_structural_equivalence(alge,
k = "elbow"))
plot(node_structural_equivalence(alge,
k = "silhouette"))
Ok, so it looks like the elbow method returns k == 3
as
a good trade-off between fit and parsimony. The silhouette method, by
contrast, sees k == 4
as maximising cluster similarity and
dissimilarity. Either is probably fine here, and there is much debate
around how to select the number of clusters anyway. However, the
silhouette method seems to do a better job of identifying how unique the
16th node is. The silhouette method is also the default in
{migraph}
.
Note that there is a somewhat hidden parameter here,
range
. Since testing across all possible numbers of
clusters can get computationally expensive (not to mention
uninterpretable) for large networks, {migraph}
only
considers up to 8 clusters by default. This however can be modified to
be higher or lower, e.g. range = 16
.
Finally, one last option is k = "strict"
, which only
assigns nodes to the same partition if there is zero distance between
them. This is quick and rigorous solution, however oftentimes this
misses the point in finding clusters of nodes that, despite some
variation, can be considered as similar on some dimension.
plot(node_structural_equivalence(alge, k = "strict"))
Here for example, no two nodes have precisely the same tie-profile,
otherwise their branches would join/fork at a distance of 0. As such,
k = "strict"
partitions the network into 16 clusters. Where
networks have a number of nodes with strictly the same profiles, such a
k-selection method might be helpful to recognise those in exactly the
same structural position, but here it essentially just reports nodes’
identity.
Blockmodelling
Summarising profiles
Ok, so now we have a result from establishing nodes’ membership in structurally equivalent classes. We can graph this of course, as above:
alge %>%
mutate(se = node_structural_equivalence(alge)) %>%
autographr(node_color = "se")
While this plot adds the structurally equivalent classes information to our earlier graph, it doesn’t really help us understand how the classes relate. That is, we might be less interested in how the individuals in the different classes relate, and more interested in how the different classes relate in aggregate.
One option that can be useful for characterising what the profile of
ties (partners) is for each position/equivalence class is to use
summary()
. It summarises some census result by a partition
(equivalence/membership) assignment. By default it takes the average of
ties (values), but this can be tweaked by assigning some other summary
statistic as FUN =
.
# Let's wrap node_tie_census inside the summary() function
# and pass it a membership result
summary(node_tie_census(____),
membership = ____)
summary(node_tie_census(alge),
membership = node_structural_equivalence(alge))
This node census produces 96 columns, \(16 \text{nodes} * 2 \text{directions} * 3 \text{edge types}\), it takes a bit to look through what varies between the different classes as ‘blocked’. But only four rows (the four structurally equivalent classes, according to the default).
Another way to do this is to plot the blockmodel as a whole. Passing
the plot()
function an adjacency/incidence matrix along
with a membership vector allows the matrix to be sorted and framed
(without the membership vector, just the adjacency/incidence matrix is
plotted):
# Let's plot the blockmodel using the plot() function we used for the dendrograms
# Instead of node_tie_census() let's us as_matrix()
plot(as_matrix(____),
membership = ____)
# plot the blockmodel for the whole network
plot(as_matrix(alge),
membership = node_structural_equivalence(alge))
# plot the blockmodel for the friends, tasks, and social networks separately
plot(as_matrix(friends),
membership = node_structural_equivalence(alge)) +
plot(as_matrix(tasks),
membership = node_structural_equivalence(alge)) +
plot(as_matrix(social),
membership = node_structural_equivalence(alge))
By passing the membership argument our structural equivalence results, the matrix is re-sorted to cluster or ‘block’ nodes from the same class together. This can help us interpret the general relationships between classes. For example, when we plot the friends, tasks, and social networks using the structural equivalence results, we might characterise them like so:
- The first group work together only in reciprocal pairs on tasks, preferring to approach the nerd but also those of the other two roles. While they hang out with each other socially quite a bit, friendship from groups 2 and 3 are preferred.
- The second group also work together only in reciprocal pairs, preferring to work collaboratively with group 1 or also the nerd. They also tend to count those from group 1 as friends, and hang out with everyone else but themselves.
- The third group will work with either some in group 1 and 3, or 2, but again prefer the nerd for task advice. They are pretty good friends with each other though, and pretty happy to socialise with everyone.
- The nerd is a loner, no friends, but everyone hangs out with them for task advice.
Reduced graph
Lastly, we can consider how classes of nodes relate to one another in a blockmodel. Let’s use the 4-cluster solution on the valued network (though binary is possible too) to create a reduced graph. A reduced graph is a transformation of a network such that the nodes are no longer the individual nodes but the groups of one or more nodes as a class, and the ties between these blocked nodes can represent the sum or average tie between these classes. Of course, this means that there can be self-ties or loops, because even if the original network was simple (not complex), any within-class ties will end up becoming loops and thus the network will be complex.
(bm <- to_blocks(alge, node_structural_equivalence(alge)))
bm <- bm %>% as_tidygraph %>%
mutate(name = c("Freaks", "Squares", "Nerds", "Geek"))
autographr(bm)