多个源顶点的subcomponent(mode =“ in”)

时间:2019-01-07 03:16:17

标签: r igraph ropensci drake-r-package

问题

igraph R包中,是否有subcomponent()和/或BFS的有效实现可以处理多个源顶点?

动机

drake R package将用户的工作流程建模为相互依赖的对象和文件的DAG。 DAG仅应包含用户的目标及其上游依存关系,因此drake使用igraph::subcomponent()来消除多余的顶点。这种方法效率低下,因为v参数必须是单个顶点,因此drake最终针对用户要构建的每个目标进行新的BFS。

编辑:2019-01-10

drake现在使用different approach that ultimately relies on sequential calls to adjacent_vertices()。该方法比较笨拙,但是the speed improvement is actually quite nice。仍然追求更优雅和精致的东西。

1 个答案:

答案 0 :(得分:2)

我认为您可以使用distances()函数来执行此操作,该函数生成节点之间的距离矩阵(无边距)。这似乎只搜索一次,并且比遍历每个顶点要快得多。

示例代码:

library(igraph)
library(microbenchmark)

# generate some random testing data
set.seed(1234)
g <- erdos.renyi.game(50, .01)

# Here we make a function that iterates 
# across the vector of IDs applying the function
# and returns a list where each element is the
# ids of the subcomponents
sc_apply <- function(g) {
  vs <- V(g)
  res <- sapply(vs, function(v){as.numeric( # to facilitate comparison
    subcomponent(g, v, mode = "in")
    )})
  res
}

# Try it for testing
t1 <- sc_apply(g)

# Here we get the matrix of node distances. Infinite distance
# implies a seperate component. We iterate through rows of
# matrix to extract the set of nodes that are connected 
sc_distmat <- function(g) {
  dmat <- distances(g, mode = "in")
  res <- apply(dmat, 1, function(row){which(is.finite(row))})
  res
}

# extract for testing
t2 <- sc_distmat(g)

# check that equal (we need to sort the 
# subcomponent list elements first to facilitate comparison)
all.equal(lapply(t1, sort), t2)
#> [1] TRUE

结果是相同的-尽管值得注意的是,如果您的图形是一个巨大的组件,而不是套用,则会向您返回矩阵而不是列表,因此您需要以稍微不同的方式进行比较。

好,现在让我们看看这是多少/是否更快:

# generate graphs of different sizes (not too big because my
# laptop is borderline antique!)
set.seed(42)
small_g <- erdos.renyi.game(10, .2)
mid_g <- erdos.renyi.game(50, .1)
big_g <- erdos.renyi.game(100, .1)

# check speed improvement
microbenchmark(sc_apply(small_g), sc_distmat(small_g))
#> Unit: microseconds
#>                 expr      min        lq      mean   median        uq
#>    sc_apply(small_g) 2181.465 2243.4895 2734.7132 2313.005 2777.7135
#>  sc_distmat(small_g)  451.333  471.8565  840.4742  521.865  598.0845
#>        max neval cld
#>   9152.262   100   b
#>  27139.262   100  a
microbenchmark(sc_apply(mid_g), sc_distmat(mid_g))
#> Unit: microseconds
#>               expr       min        lq       mean    median         uq
#>    sc_apply(mid_g) 11006.113 11327.794 13590.9536 12919.492 15397.2510
#>  sc_distmat(mid_g)   757.752   795.308   949.2551   841.834   965.4545
#>        max neval cld
#>  27068.296   100   b
#>   2061.824   100  a
microbenchmark(sc_apply(big_g), sc_distmat(big_g))
#> Unit: milliseconds
#>               expr      min        lq      mean    median        uq
#>    sc_apply(big_g) 23.11678 26.696373 29.940675 29.191045 33.012796
#>  sc_distmat(big_g)  1.67531  1.727873  2.156307  1.855994  2.244872
#>        max neval cld
#>  47.081647   100   b
#>   7.576123   100  a

您可以看到distances()方法更快,并且随着图形大小的增长而越来越快。

reprex package(v0.2.1)于2019-01-10创建