Correlation network
I came up with an idea to draw correlation network to get a grasp about relationship between a list of stocks. An alternative way to show correlation matrix would be head map, which can have limitations with big matrices (>100).
Unfortunately, ggplot2 package doesn’t have a easy way to draw the networks, so I was left with igraph or network. I tried both, but somehow chose igraph. If you want to master either package I highly recommend to start from theoretical part – it is very well written and it will save your time trying to understand the package’s conception.
Here is correlation matrix of stocks, which correlation coef. is more than 0.5:
require(xts) require(quantmod) require(igraph) cor_mat<- matrix( runif(100), nr=10 ) cor_mat[ lower.tri(cor_mat, diag=TRUE) ]<- 0 cor_mat[ abs(cor_mat) < 0.5]<- 0 graph <- graph.adjacency(cor_mat>0.5, weighted=TRUE, mode="upper") E(graph)$weight<-t(cor_mat)[abs(t(cor_mat))>0.5] E(graph)[ weight>0.7 ]$color <- "black" E(graph)[ weight>=0.65 & weight<0.7 ]$color <- "red" E(graph)[ weight>=0.6 &weight<0.65 ]$color <- "green" E(graph)[ weight>=0.55 &weight<0.6 ]$color <- "blue" E(graph)[ weight<0.55 ]$color <- "yellow" V(graph)$label<- seq(1:10)#V(graph)$name graph$layout <- layout.fruchterman.reingold factor<-as.factor(cut(E(graph)$weight*10,c(4,5,6,7,8),labels=c(1,10,20,30))) png('corr_network.png',width=500) plot(decompose.graph(graph)[[which.max(sapply(decompose.graph(graph), vcount))]],edge.width =as.numeric(factor)*1.5,frame=T) legend("bottomleft", title="Colors", cex=0.75, pch=16, col=c("black", "blue","red", "green","pink"), legend=c(">70%", "65-70","60-65","55-60","50-55"), ncol=2) dev.off() |
The code can be found on github.
Justas said,
March 22, 2011 @ 19:57
How many days/months did you take for calculating correlations?
Dzidorius Martinaitis said,
March 22, 2011 @ 20:18
60 working days
Tal Galili said,
March 23, 2011 @ 7:25
Great idea!
I wonder how it might be improved…
Best,
Tal
Justas said,
March 23, 2011 @ 9:19
Well, it might be improved by using some other visualization tools to see how correlations change over time. For instance, start from Year X and and end at Year Z and we would see how relationships vary.
Jens said,
March 23, 2011 @ 16:50
Thank you, nice idea!
I adapted this immediately to an analysis of couple behaviours and attitudes.One point that gave me a headache was labeling the vertices, that part was missing from your code example. I accomplished this by (simplified code snippet) :
g <- decompose.graph(graph)[[which.max(sapply(decompose.graph(graph), vcount))]]
plot(g, vertex.label=V(g)$name)
Regards, jens
Dzidorius Martinaitis said,
March 23, 2011 @ 17:01
Jens,
it is hidden on this line:
V(graph)$label<- seq(1:10)#V(graph)$name
For my personal use I generated correlation matrix from the real data, but then I wrote this post, I generated random matrix and assigned the names like seq(1:10). If you want the nice labels on vertices, you can do your way or this way:
V(graph)$label<-V(graph)$name
Thanks for pointing out the problems!
DavidC said,
March 24, 2011 @ 3:35
This is great. Thanks!
Tal Galili said,
March 24, 2011 @ 8:47
Hi there,
I just found a way for plotting a network in ggplot2:
http://www.r-bloggers.com/basic-ggplot2-network-graphs-ver2/
I’d be glad to see you use it for the correlation networks
Cheers,
Tal
Sacha Epskamp said,
March 26, 2011 @ 14:36
You might also be interested in my package which is focussed around doing just this
https://sites.google.com/site/qgraphproject/examples
zubin said,
April 24, 2011 @ 23:08
the qgraph package referenced above is amazing. just tried it out, correlation networks are extremely intuitive.