[This article was first published on Peter Solymos - R related posts, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

The mefa4 R package is aimed at efficient manipulation of very big data sets leveraging sparse matrices thanks to the Matrix package. The recent update (version 0.3-3) of the package includes a bugfix and few new functions to compare sets and finding dominant features in compositional data as described in the ChangeLog.

The first new function is compare_sets. It comes really handy when one needs to compare row names for two matrix like objects. Such as when trying to merge two tables which come from different sources. This facilitates checking the data and troubleshooting. The function takes two arguments which are then compared both in terms of unique values, and in terms of levels when input is a factor (referred to as labels). The function combines the functionality of length (as in length(unique(...))), nlevels, union, intersect, and setdiff. In the first example let us compare two numeric vectors:

compare_sets(x = 1:10, y = 8:15)
##        xlength ylength intersect union xbutnoty ybutnotx
## labels      10       8         3    15        7        5
## unique      10       8         3    15        7        5


Now let us have a look at two factors, one with ‘zombie’/empty/unused levels. In this case the two rows differ for obvious reasons:

compare_sets(x = factor(1:10, levels=1:10), y = factor(8:15, levels=1:15))
##        xlength ylength intersect union xbutnoty ybutnotx
## labels      10      15        10    15        0        5
## unique      10       8         3    15        7        5


The second function is called find_max. No, it is not a dog locator, and it has nothing to do with Ruby. It takes a matrix-like object as its argument and finds the maximum value and column ID for each row. Such a function is handy when for example one is looking for a dominant feature type in a matrix of compositional data. For example the area of discrete habitats is summarized in buffers around point locations using some GIS application. As a result, we have a matrix where rows sum to 1 (note: this is not a criteria for the function to work):

mat <- matrix(runif(10 * 5, 0, 1), 10, 5)
set.seed(1)
mat <- matrix(runif(10 * 5, 0, 1), 10, 5)
mat <- mat / rowSums(mat)
colnames(mat) <- paste0("V", 1:5)
mat
round(mat, 3)
##          V1    V2    V3    V4    V5
##  [1,] 0.098 0.076 0.345 0.178 0.303
##  [2,] 0.185 0.088 0.106 0.299 0.322
##  [3,] 0.180 0.216 0.204 0.155 0.246
##  [4,] 0.421 0.178 0.058 0.086 0.256
##  [5,] 0.078 0.297 0.103 0.319 0.204
##  [6,] 0.277 0.154 0.119 0.206 0.244
##  [7,] 0.379 0.288 0.005 0.319 0.009
##  [8,] 0.252 0.379 0.146 0.041 0.182
##  [9,] 0.189 0.114 0.261 0.217 0.220
## [10,] 0.027 0.340 0.149 0.180 0.303


The find_max function output has an index column with the column names where values were maximum, and the value itself. Numeric column indices can be recovered by coercing the values in the index column to integers:

(m <- find_max(mat))
##    |++++++++++++++++++++++++++++++++++++++++++++++++++| 100% ~00s
##    index     value
## 1     V3 0.3450096
## 2     V5 0.3223295
## 3     V5 0.2455856
## 4     V1 0.4210278
## 5     V4 0.3187309
## 6     V1 0.2772785
## 7     V1 0.3788923
## 8     V2 0.3785517
## 9     V3 0.2607874
## 10    V2 0.3404492
as.integer(m\$index)
##  [1] 3 5 5 1 4 1 1 2 3 2


You might wonder what the third function find_min might do. Install the package from your nearest CRAN mirror by install.packages("mefa4") to find out! Let me know if you find these updates and the package useful, or have feature requests, find issues etc. on the GitHub development site.

A next post will tell more about what that ~00s at the end of the progress is all about (hint: the pbapply package is now a dependency).