Description
In this assignment, we will study a real network. The data is available on this Dropbox Link 1
. We are going to study the properties of
this real network. The data is given as a directed edge list format,
where each line has three items: node1, node2 and the weight of the
edge from node1 to node2.
1. Is this network connected? If not, find out the giant connected
component. And in the following, we will deal with this giant
connected component.
2. Measure the degree distribution of in-degree and out-degree of
the nodes.
3. We would like to measure the community structure of the network. First, we need to convert it into an undirected network.
Then we use fastgreedy.community and label.propagation.community
to measure the community structure. Are the results of these two
methods similar or not?
1https://www.dropbox.com/s/oybns7fvztfy76q/sorted_directed_
net.txt?dl=0
1
Note that this network has edge weights, so it is not trivial to
convert it from directed to undirected, especially if there are two
directed edges between node i and node j. We at least have two
options.
In option 1, we keep the number of edges unchanged, and just
remove the directions. The resulting undirected network is not
simple. label.propagation.community command can be used
to compute a weighted, undirected, non-simple network’s community structure.
In option 2, we merge the two directed edges between i and j so
that the resulting network is simple. Suppose the weights are wij
and wji respectively, and we set the weight for the merged undirected edge to be √wijwji. We can use both fastgreedy.community
and label.propagation.community to measure the community structure.
4. Find the largest community computed from fastgreedy.community.
Isolate the community from other parts of the network to form a
new network, and then find the community structure of this new
network. This is the sub-community structure of the largest community.
5. Find all the sub-community structures of the communities with
sizes larger than 100.
6. Both fastgreedy.community and label.propagation.community
assume that each node belongs to only one community. In practice, a node can belong to two or more communities at the same
time. There is no command in igraph that can detect overlapped
communities. Here we are going to use personalized PageRank to
study the overlapped communities structures.
Suppose that we start the random walk from a node i with the
damping parameter 0.85 in the original directed network. We will
obtain the visiting probabilities of all nodes in the network. These
visiting probabilities reflect the relation between the other nodes
and node i. On the other hand, we already have some knowledge of the community structure obtained from the above ques2
tion. Therefore, we can compute
M~
i =
X
j
vj ~mj
,
where vj is the visiting probability of node j and ~mj is its community membership computed from 3. Here, the community membership is expressed as a vector so that it can be adapted to multiple memberships. Suppose you have found n communities in 3,
and ~mj is a n dimensional vector with only one element being 1.
The resulting M~
i is a measurement of the multiple memberships.
To include all nodes in the network in the sum might make the
computation amount too large. So you can take the largest 30 vj
to include in the sum.
A threshold is needed to remove the memberships that have very
small values in M~
i
. Choose a proper threshold.
Note that visit probabilities of the random walk started from a
node can be obtained by a personalized pagerank with a teleportation probability distribution that is 1 to the node under examination and 0 to every other node.
Try this method to see whether you can find examples of nodes
that belong to multiple communities. List at least 3 examples if
you find such nodes.
3