[R] How to pre-filter large amounts of data effectively

[Torsten Schindler]

Hi,

I'm a R newbie and want to accelerate the following pre-filtering  
step of a data set with more than 115,000 rows :

#-----------------
# Function to filter out constant data columns
filter.const<-function(X, vectors=c('column', 'row'), tol=0){
   realdata=c()
   filteredX<-matrix()
   if( vectors[1] == 'row' ){
     for( row in (1:nrow(X)) ){
       if( length(which(X[row,]!=median(X[row,])))>tol ){
         realdata[length(realdata)+1]=row
       }
     }
     filteredX=X[realdata,]
   } else if( vectors[1] == 'column' ){
     for( col in (1:ncol(X)) ){
       if( length(which(X[,col]!=median(X[,col])))>tol ){
         realdata[length(realdata)+1]=col
       }
     }
     filteredX=X[,realdata]
   }
   return(list(x=filteredX, ix=realdata))
}

#-----------------
# Filter out all all-constant columns in my training data set
#
# Read training data set with class information in the first column
training <- read.csv('training_data.txt')
dim(training) # => 49 rows and 525 columns

# Prepare column names by stripping the underline and the number at  
the end
colnames(training) <- sub('_\\d+$', '', colnames(training), perl=TRUE)

# Filter out the all-constant columns, exclude column 1, the class  
column called myclass
training.filter <- filter.const(training[,-1])

# The filtered data frame is
training.filtered <- cbind(myclass=training[,1], training.filter$x)
dim(training.filtered) # => 49 rows and 250 columns

# Save the filtered training set for later use in classification
filtered.data <- 'training_set_filtered.Rdata'
save(training.filtered, file=filtered.data)

#-----------------
# THE FOLLOWING FILTERING STEP TAKES 3 HOUR ON MY PowerBook
# AND CONSUMES ABOUT 600 Mb MEMORY.
#
# I WOULD BE HAPPY ABOUT ANY HINT HOW TO IMPROVE THIS.

# Pre-filter the big data set (more than 115,000 rows and 524  
columns) for later class predictions.
# The big data set contains the same column names as the training  
set, but in a different order.

input.file <- 'big_data_set.txt'
filtered.file <- 'big_data_set_filtered.txt'

# Read header with first row
prediction.set <- read.csv(input.file, header=TRUE, skip=0, nrow=1)

# Prepare column names by stripping the underline and the number at  
the end
colnames(prediction.set) <- sub('_\\d+$', '', colnames 
(prediction.set), perl=TRUE)
prediction.set.header <- colnames(prediction.set)

# Get descriptor columns of the training data set without the  
Activity_Class column
training.filtered.property.colnames <- colnames(training.filtered)[-1]

# Filter out the all-constant columns from the training set
prediction.set.filtered <- prediction.set 
[training.filtered.property.colnames]
dim(prediction.set.filtered) # => 1 row and 249 columns

# Write header and the first filtered row
write.csv(prediction.set.filtered, file=filtered.file,
             append=FALSE,  
col.names=training.filtered.property.colnames)

blocksize <- 1000
for (lineid in (0:120)*blocksize) {
   cat('lineid: ', lineid, '\n')

   # Read block of data
   # We have to add an dummy colname "x" in the col.names, when the  
header is not read!
   prediction.set <- try(read.csv(input.file, header=FALSE,
                         col.names=c('x',prediction.set.header),  
row.names=1,
                         skip=lineid+2, nrow=blocksize))
   if (class(prediction.set) == "try-error") break

   # Filter out all-constant training set columns from the block
   prediction.set.filtered <- prediction.set 
[training.filtered.property.colnames]

   # Append the data
   # (I know this function is slow, but I couldn't figure out how to  
do it faster, so far.)
   write.table(prediction.set.filtered, file=filtered.file,
                         append=TRUE, col.names=FALSE, sep=",")
}

#-------------
# Now read in the filtered data set and save it for later use in  
classification
prediction.set.filtered <- read.csv(filtered.file, header=TRUE,  
row.names=1)
filtered.data <- 'prediction_set_filtered.Rdata'
save(prediction.set.filtered, file=filtered.data)



I would be very happy about any hints how to improve the code above!!!

Best regards,

Torsten