Structuring data for community detection in R's igraph

Question

I am new to igraph and social network analysis, but not to R.

I am struggling to correctly structure my dataset for community detection, but have successfully used iGraph to generate a co-occurence matrix as directed [here]. What I would like to do next is use a community detection algorithm on the same dataset to create a graph showing clusters as is done in the answer here.

The sample code for how to do this is as follows:

df1 <- graph.famous("Zachary")
df2 <- walktrap.community(df1) #any algorithm
plot.communities(df2, df)

I've been poking around on the web to find out the structure of the Zachary dataset so I can correctly model my data, but am struggling to find my way through the technical documentation.

My data is currently structured in long form, such that:

id         interest    comments
1             Comedy          2
1  Music: Electronic         11
1       Video Gaming         10
1         Music: Pop          1
1      Entertainment          1
1       Video Gaming          4
2       Video Gaming         45
2      Entertainment         26
2         Music: Pop          1
2            Comedy         14
3      Video Gaming         10
3     Entertainment          4
3            Comedy          8
4      Video Gaming          9
4 Music: Electronic          1
4        Music: Pop          2
5        Music: Pop          2
5     Entertainment          1
5      Video Gaming          1
6      Video Gaming         12

I am trying to find clusters of overlapping interest in the population I am studying, so the ID is a person, the interests are the person's interests, and comments is an index of how many times they have shown interest. Does this help?

I've tried to run the community algorithms on this dataset (e.g. df2 <- walktrap.community(df)) but that doesn't seem to work correctly. Thoughts on what this n00b is doing wrong?

Answer 1

A graph won't allow you to cluster individuals, but only to find out which variables are related. Still, if you want to build a graph from your data, here is what you would have to do. (Note that I have saved your sample data as a .csv file and uploaded it to dropbox to make for a readily reproducible example).

library(repmis)
library(igraph)

test=source_data("https://www.dropbox.com/s/bochkedd4o3gzvq/so.csv?dl=0")

First, what you want is to create a matrix with one row per individual and one column per feature:

matrix=matrix    
rownames(matrix)=unique(test[,1])
colnames(matrix)=unique(test[,2])

and where the values of the cells correspond to the strength of the interests:

for (i in 1:nrow(matrix)){
    temp=test[test[,1]==i,][,2:3]
    for (j in 1:ncol(matrix)){
    matrix[i,j]=sum(temp[temp[,1]==colnames(matrix)[j],2]) # sum is used because of duplicates
    }
}

What you get is:

> matrix
  comedy electronic gaming pop ent
1      2         11     14   1   1
2     14          0     45   1  26
3      8          0     10   0   4
4      0          1      9   2   0
5      0          0      1   2   1
6      0          0     12   0   0

Then, from that, you can create an adjacency matrix:

x=t(matrix)%*%matrix

And what you get is:

> x
           comedy electronic gaming pop  ent
comedy        264         22    738  16  398
electronic     22        122    163  13   11
gaming        738        163   2547  79 1225
pop            16         13     79  10   29
ent           398         11   1225  29  694

Building a graph from that is easy:

g=graph.adjacency(x,weighted=T,mode="undirected",diag=FALSE)
g=simplify(g)

You can apply any community detection algorithm to the object g, for instance:

spinglass.community(g,weight=E(g)$weight)

If you want to cluster individuals rather than variables, I would look at PCA and hierarchical clustering (see for instance the excellent HCPC function from the FactoMineR package). You will use in that case the object matrix above (no need to compute the adjacency matrix).