Transaction problem in RStudio for tweet apriori analysis - r

Question

I want to use the apriori algorithm to apply association rules between words on the tweet database I have with RStudio. However, the code below gives an error on a million rows of data, while working on a small number of data. I needed your help as I couldn't understand what caused the error.
TweetTrans <- read.transactions("../input/tweets/output.csv",
rm.duplicates=FALSE,
format = "basket",
sep = ",",
encoding = "UTF-8")
The Error is:
Error in validObject(.Object): invalid class “ngCMatrix” object: row indices are not sorted within columns
Traceback:
1. read.transactions("../input/tweets/output.csv", rm.duplicates = FALSE,
. format = "basket", sep = ",", encoding = "UTF-8")
2. as(data, "transactions")
3. asMethod(object)
4. new("transactions", as(from, "itemMatrix"), itemsetInfo = data.frame(transactionID = names(from),
. stringsAsFactors = FALSE))
5. initialize(value, ...)
6. initialize(value, ...)
7. callNextMethod()
8. .nextMethod(.Object = .Object, ... = ...)
9. callNextMethod()
10. .nextMethod(.Object = .Object, ... = ...)
11. as(from, "itemMatrix")
12. asMethod(object)
13. new("ngCMatrix", p = c(0L, p), i = as.integer(i) - 1L, Dim = c(length(levels(i)),
. length(p)))
14. initialize(value, ...)
15. initialize(value, ...)
16. callNextMethod()
17. .nextMethod(.Object = .Object, ... = ...)
18. validObject(.Object)
19. stop(msg, ": ", errors, domain = NA)

Answer 1

Here are some ideas for how to find a rogue line in the data file. The input to read.transactions should be a text file the looks something like
A, B, C
B, C
C, D, E
D, A, B, F
where A, B ,C, etc are the names of the items (probably longer than one character each!)
So you could read in the file using readLines...
data <- readLines("../input/tweets/output.csv")
Each element of data (one per line of the file) should be a string of the form "A, B, C" etc, as above.
You could then use functions (e.g. from the stringr package) to check if any lines contain unusual characters, or have an odd format. Without seeing your file, it is hard to say how to do this, but you might, for example, look for quotes in odd places (str_detect(data, '\\"')) or characters that are not letters, digits , spaces or commas (str_detect(data, "[^\\w\\d\\s,]")).
Another thing you could try is to write a for loop to take each element of data (or perhaps larger chunks if that is too slow), save it as a file, try reading it with read.transactions, and see where it crashes.
for(i in seq_along(data)){
writeLines(data[i], "dummyfile.csv")
trans <- read.transactions("dummyfile.csv",
rm.duplicates=FALSE,
format = "basket",
sep = ",",
encoding = "UTF-8")
}
The value of i when it crashes will give you the problem row number. It might take a long time to run, though!

Answer 2

I ran into a very similar problem: the same error got triggered when trying to cast a list to a transaction object.
I also couldn't easily figure out what lines in the data caused the issue, as it seems to be triggered by a combination of transactions and not necessarily by any individual one, but I managed to track down the source of the problem in this assignment (source):
p <- new("ngCMatrix", p = c(0L, p),
i = as.integer(i) - 1L,
Dim = c(length(levels(i)), length(p)))
My R got pretty rusty over time and I couldn't find an immediate way to patch the code, but I came up with an alternative solution for constructing the ngCMatrix object:
Assume you have the data in a data.frame following some sort of (user, item) format - in your case it would most likely be (tweet_id, term/word)
Create a unique incremental ID for every user and item and add it to your data.frame
Use those ID to create the sparse matrix and - optionally - enrich it with the labels for item and user to make it more interpretable
Finally, cast the sparse matrix to a transaction object
Example (I implemented mine with data.table, but a traditional dataframe implementation would be very similar):
library(Matrix)
library(data.table)
library(arules)
DT <- data.table(user = c('A','A','B','B','A','C','D'),
item = c('AAB','AAA','AAB','BBB','ABA','BBB','AAB'))
# Create user_ids
unique_users <- unique(DT$user)
users <- data.table(user=unique_users,
user_id=c(1:length(unique_users)))
# Repeat for items
unique_items <- unique(DT$item)
items <- data.table(item=unique_items,
item_id=c(1:length(unique_items)))
# Add indexes to original data table (setting keys helps with performance)
DT <- merge.data.table(x=DT, y=users, by='user')
DT <- merge.data.table(x=DT, y=items, by='item')
# Create the sparse matrix
mat <- sparseMatrix(
i = DT$item_id,
j = DT$user_id,
dims = c(nrow(items), nrow(users)),
dimnames = list(items$item, users$user)
)
# transform to arules 'transactions'
txn <- as(op, "transactions")
Please note that this doesn't help understanding what caused the issue, but rather provides a workaround to solve it. In my data.table implementation the code is pretty performant, taking only a few seconds to process over 30M transactions on a laptop-sized machine (2 CPUs, 16gb RAM).