Part of a project for Network analysis class is constructing package dependency tree off of CRAN (I called shotgun). I could scrape CRAN website for packages and their dependencies, suggests, imports and enhancements (as done here), but I opted for a more elegant solution via available.packages function, that reports which packages are available on CRAN and also adds all the above relationships. Feel free to add to this function by pulling my GitHub repo.
library(igraph)
library(reshape2)
library(stringr)
library(RColorBrewer)
# Get a list of available packages.
dep.data <- available.packages(contriburl = "http://cran.at.r-project.org/bin/windows/contrib/3.0")
dep.data[1:5, 1:5]
## Package Version Priority Depends
## A3 "A3" "0.9.2" NA "xtable, pbapply"
## ABCExtremes "ABCExtremes" "1.0" NA "SpatialExtremes, combinat"
## ABCp2 "ABCp2" "1.0" NA "MASS"
## ACCLMA "ACCLMA" "1.0" NA NA
## ACD "ACD" "1.5" NA "R(>= 2.13.0)"
## Imports
## A3 NA
## ABCExtremes NA
## ABCp2 NA
## ACCLMA NA
## ACD NA
# Get a subset of variables of interest.
dep.data <- dep.data[, c("Package", "Depends", "Imports", "Suggests", "Enhances")]
# Some (most?) packages depends/import to multiple packages. Let's tease
# those packages apart. This function will make a 'long' list of
# relationships for each package (where applicable). If there's no
# relationship, the output is 'none'.
make.data <- apply(dep.data, 1, FUN = function(x) {
if (all(is.na(x[-1])))
return(data.frame(from = x[1], to = x[1], relation = "none"))
x2 <- na.omit(x[-1])
x2 <- sapply(x2, strsplit, ",")
out <- data.frame(from = rep(x[1], length(unlist(x2))), to = unlist(x2),
relation = rep(names(x2), times = sapply(x2, length)))
rownames(out) <- 1:nrow(out)
out
})
done.data <- do.call("rbind", make.data)
rownames(done.data) <- 1:nrow(done.data)
# We clean the data.
clean.data <- done.data # make a copy
# remove version numbers
clean.data$to <- sub("\\((.*?)\\)", "", x = clean.data$to, perl = TRUE)
# remove newlines
clean.data$to <- sub("\\\n", "", x = clean.data$to, perl = TRUE)
# remove dependency on R (*)
clean.data[clean.data$to == "R" | clean.data$to == "R ", "to"] <- as.character(clean.data[clean.data$to ==
"R" | clean.data$to == "R ", "from"])
# clean whitespace
clean.data$to <- as.factor(str_trim(clean.data$to, "both"))
# remove loops
clean.data <- clean.data[!apply(clean.data[, c("from", "to")], 1, function(x) x[1] ==
x[2]), ]
This is the function that will be doing the heavy drawing for us.
#' Draws a subset of a data.frame as a network graph.
#' @param x data.frame. Has at least three columns, from, to and relation.
#' @param relation Character vector of length one. Rows with what relation from column relation (see x) should be used?
#' @param pkg Character. Package name to be subsetted.
#' @param ... Arguments passed to plot function that plots igraph object.
#' @return \code{igraph} object with the possible side effect of plotting the object.
plotGraph <- function(x, relation = NULL, pkg.to = NULL, pkg.from = NULL, plot = TRUE,
...) {
# prepare the data to be plotted
if (!is.null(relation)) {
gd <- droplevels(x[x$relation %in% relation, ])
} else {
gd <- x
}
if (!is.null(pkg.to))
gd <- gd[gd$to %in% pkg.to, ]
if (!is.null(pkg.from))
gd <- gd[gd$from %in% pkg.from, ]
# make the data an igraph object and plot it (if specified)
g <- graph.data.frame(gd, directed = TRUE)
cc <- E(g)$relation
if (any(unique(cc) == "Imports"))
cc[cc == "Imports"] <- "#E41A1C"
if (any(unique(cc) == "Suggests"))
cc[which(cc == "Suggests")] <- "#984EA3"
if (any(unique(cc) == "Enhances"))
cc[which(cc == "Enhances")] <- "#4DAF4A"
if (any(unique(cc) == "Depends"))
cc[which(cc == "Depends")] <- "#377EB8"
E(g)$color <- cc
if (plot)
plot(g, ...)
# see ?igraph:::layout for possible layout options
g
}
We can try and draw a full dependency graph, but it doesn't look pretty. It also takes a long time. Consider before executing this next chunk. Output of plotGraph function is an igraph object, which is why we save it to a variable for possible further manipulation.
graph.all.depends <- plotGraph(x = clean.data, relation = "Depends", vertex.label.color = "blue",
edge.width = 0.5, edge.color = "#ACC49D", edge.arrow.size = 0.5, vertex.label.cex = 0.7,
vertex.size = 0, vertex.color = "white", layout = layout.kamada.kawai)
Another example of usage would be to plot the relationships between packages for high performance computing (see HPC task view on CRAN).
hpp <- c("batch", "BatchExperiments", "BatchJobs", "bayesm", "bcp", "biglars",
"biglm", "bigmemory", "bigrf", "biopara", "bnlearn", "caret", "cudaBayesreg",
"data.table", "dclone", "doMC", "doMPI", "doRedis", "doRNG", "doSNOW", "ff",
"foreach", "fork", "GAMBoost", "gcbd", "Geneland", "gputools", "GridR",
"HadoopStreaming", "harvestr", "HiPLARM", "inline", "latentnet", "lga",
"magma", "mapReduce", "Matching", "mchof", "multicore", "nws", "OpenCL",
"orloca", "partDSA", "pbdBASE", "pbdDEMO", "pbdDMAT", "pbdMPI", "pbdNCDF4",
"pbdSLAP", "peperr", "permGPU", "pmclust", "profr", "proftools", "pvclust",
"Rcpp", "Rdsm", "rgenoud", "RInside", "rJava", "rlecuyer", "Rmpi", "rpud",
"rredis", "rsprng", "scaleboot", "snow", "snowfall", "snowFT", "speedglm",
"sprint", "sqldf", "STAR", "tm", "varSelRF", "WideLM", "xgrid")
graph.parallel.depends <- plotGraph(x = clean.data, pkg.to = hpp, relation = "Depends",
vertex.label.color = "black", edge.width = 1, edge.arrow.size = 0.2, vertex.label.cex = 0.9,
vertex.size = 0, vertex.color = "white", layout = layout.kamada.kawai)
Or perhaps we'd be interested in seeing relationships for parallel computing packages. We notice that foreach, snow and Rmpi are quite popular. multicore and snowfall (this one being my fav) have less relationships.
Oh, I forgot to mention, lines are color coded and denote different relationship types (colors were taken from brewer.pal(4, "Set1")).
graph.somepara.all <- plotGraph(x = clean.data, pkg.to = c("snowfall", "snow",
"foreach", "Rmpi", "multicore", "sprint", "biopara", "bigrf", "doMPI", "doRedis",
"nws", "snowFT"), vertex.label.color = "black", edge.width = 1, edge.arrow.size = 0.2,
vertex.label.cex = 0.9, vertex.size = 0, vertex.color = "white", layout = layout.kamada.kawai)
Let's see only depends for parallel computing packages. Notice that depends are colored blue(ish). Quite a few packages depend on foreach.
graph.somepara.depends <- plotGraph(x = clean.data, pkg.to = c("snowfall", "snow",
"foreach", "Rmpi", "multicore", "sprint", "biopara", "bigrf", "doMPI", "doRedis",
"nws", "snowFT"), relation = "Depends", vertex.label.color = "black", edge.width = 1,
edge.arrow.size = 0.2, vertex.label.cex = 0.9, vertex.size = 0, vertex.color = "white",
layout = layout.kamada.kawai)
I also fancy packages for spatial data. Barry has made a contribution on this subject, and you can find the graphs here.
A lot of packages depend on sp.
graph.sp2 <- plotGraph(x = clean.data, pkg.to = "sp", relation = "Depends",
vertex.label.color = "black", edge.width = 0.1, edge.color = "#377EB8",
vertex.label.cex = 0.9, vertex.size = 0, vertex.color = "white", layout = layout.kamada.kawai)
spp <- c("ade4", "adehabitat", "adehabitatHR", "adehabitatHS", "adehabitatLT",
"adehabitatMA", "ads", "akima", "ash", "aspace", "automap", "classInt",
"CompRandFld", "constrainedKriging", "cshapes", "DCluster", "deldir", "DSpat",
"ecespa", "fields", "FieldSim", "gdistance", "Geneland", "GEOmap", "geomapdata",
"geonames", "geoR", "geoRglm", "geosphere", "geospt", "geospt", "GeoXp",
"ggmap", "glmmBUGS", "gmt", "GriegSmith", "gstat", "Guerry", "hdeco", "intamap",
"landsat", "mapdata", "mapproj", "maps", "maptools", "MarkedPointProcess",
"MBA", "Metadata", "micromap", "ModelMap", "ncdf", "ncf", "nlme", "OpenStreetMap",
"osmar", "pastecs", "PBSmapping", "PBSmodelling", "plotKML", "psgp", "ramps",
"RandomFields", "rangeMapper", "RArcInfo", "raster", "rasterVis", "RColorBrewer",
"regress", "rgdal", "rgeos", "RgoogleMaps", "RPyGeo", "RSAGA", "RSurvey",
"rworldmap", "sgeostat", "shapefiles", "sp", "spacetime", "sparr", "spatcounts",
"spatgraphs", "spatial", "spatialCovariance", "SpatialExtremes", "spatialkernel",
"spatialprobit", "spatialsegregation", "spatstat", "spBayes", "spcosa",
"spdep", "spgrass6", "spgwr", "sphet", "splancs", "spsurvey", "Stem", "tgp",
"trip", "tripack", "tripEstimation", "UScensus2000blkgrp", "UScensus2000cdp",
"UScensus2000tract", "vardiag", "vec2dtransf", "vegan")
graph.spp <- plotGraph(x = clean.data, pkg.to = spp, relation = "Depends", vertex.label.color = "black",
edge.width = 0.1, edge.color = "#377EB8", vertex.label.cex = 0.5, vertex.size = 0,
edge.arrow.size = 0.2, vertex.color = "white", layout = layout.kamada.kawai)
What a mess! Let's see what's going on for the top few packages.
top.sp <- names(sort(degree(graph.spp), decreasing = TRUE)[1:20])
graph.top.spp <- plotGraph(x = clean.data, pkg.to = top.sp, relation = "Depends",
vertex.label.color = "black", edge.width = 0.1, edge.color = "#377EB8",
vertex.label.cex = 0.5, vertex.size = 0, edge.arrow.size = 0.2, vertex.color = "white",
layout = layout.kamada.kawai)
Better, but let's narrow it down even further. We exclude nlme and RColorBrewer.
top.sp <- names(sort(degree(graph.spp), decreasing = TRUE)[1:12])
top.sp <- top.sp[which(!top.sp %in% c("nlme", "RColorBrewer"))]
graph.top.spp <- plotGraph(x = clean.data, pkg.to = top.sp, relation = "Depends",
vertex.label.color = "black", edge.width = 0.1, edge.color = "#377EB8",
vertex.label.cex = 0.6, vertex.size = 0, edge.arrow.size = 0.2, vertex.color = "white",
layout = layout.kamada.kawai)
It's up to you to explore your favorite packages' networks further. For comments or questions, feel free to contact me.