I use to store my sparse data in MariaDB table with DynamicColumns structure.
When I am importing it to R I'm getting the following db_frame data frame:
db_frame <- dbGetQuery(mydb, "SELECT uid, column_json(groups) FROM matrix")
db_frame
uid column_json(groups)
1 8 {"33755311":1,"58534882":1}
2 9 {"75338985":1}
3 15 {"5445504":1,"58534882":1}
4 16 {"14897324":1,"22522055":1,"68471405":1}
5 20 {"22522055":1,"48940689":1}
My aim is to convert it from the above Compressed Row Storage to the following sparse-format(!) Matrix:
14897324 22522055 33755311 48940689 5445504 58534882 68471405 75338985
8 . . 1 . . 1 . .
9 . . . . . . . 1
15 . . . . 1 1 . .
16 1 1 . . . . 1 .
20 . 1 . 1 . . . .
I have done this in two ways but it seems to me both are very inefficient. In real life number of variables (columns) is about 2-3K, and number of observations (rows) ~ 1 million. So my both methods take a very long time to finish the work (days).
Is there other, more elegant way to do this conversion? Thank you in advance!
UPD: There are links to two dataframes in R format:
UPD2: SPEED COMPARISON TABLE for convertion data frame with 10K observation
Core i3 2.93 Ghz
Method1 My 162-188 sec Method2 My 94-102 sec Method3 @amatsuo_net 47-57 sec (but not exactly required format) Method4 @amatsuo_net+My 6 sec
For your information my two methods:
The first step is to convert db_frame data frame to more friendly list to have a possibility to extract variables names for each observation from it.
library("rjson")
var_list <- lapply(db_frame[,-1],fromJSON)
var_list_names <- lapply(var_list,names)
var_list_names
[[1]]
[1] "33755311" "58534882"
[[2]]
[1] "75338985"
[[3]]
[1] "5445504" "58534882"
[[4]]
[1] "14897324" "22522055" "68471405"
[[5]]
[1] "22522055" "48940689"
Lists of variables and observations names:
groups <- sort(unique(unlist(var_list_names)))
groups
[1] "14897324" "22522055" "33755311" "48940689" "5445504" "58534882" "68471405" "75338985"
uids <- db_frame$uid
uids
[1] "8" "9" "15" "16" "20"
------------- Method 1 using "for i" cycle
row_number = length(uids)
col_number = length(groups)
# creating empty sparse matrix M1
M1 <- sparseMatrix(dims = c(row_number,col_number), i={}, j={}, x=1)
rownames(M1) <- uids
colnames(M1) <- groups
# filling M1
for (i in 1:row_number) {
M1[i,var_list_names[[i]]] <-1
}
M1
-------------- Method 2 using "reshape2"
library("reshape2")
long <- melt(var_list)
long
value L2 L1
1 1 33755311 1
2 1 58534882 1
3 1 75338985 2
4 1 5445504 3
5 1 58534882 3
6 1 14897324 4
7 1 22522055 4
8 1 68471405 4
9 1 22522055 5
10 1 48940689 5
i=long$L1
j=match(long[,"L2"],groups)
M2 <-sparseMatrix(i=i, j=j, x=1)
rownames(M2) <- uids
colnames(M2) <- groups
M2
Thanks to @amatsuo_net who point me to rbindlist function from data.table package.
I have slightly simplified his code and added conversion to sparse format.
Conversion time on test 10K observation is impressive 6 sec!
-------------- Method 4 using "rbindlist"
library(RMySQL)
library(Matrix)
library(rjson)
library(data.table)
library(magrittr)
df <- dbGetQuery(mydb, "SELECT uid, column_json(groups) FROM matrix")
# "rbindlist" does all the work
M3 <- lapply(df[,-1],fromJSON) %>% rbindlist(fill=TRUE)
# replace NA with 0 (required for sparsematrix type)
M3[is.na(M3)] <- 0
# converting to sparsematrix type
M3 <- as(as.matrix(M3), "sparseMatrix")
# make some order :)
M3 <- M3[, order(as.integer(colnames(M3)))]
row.names(M3) <- df$uid