faster TFIDF

[jamespeapen]

Fix documentation to say this is done on counts not on compartment calls
Should this be just counts or is there a better word?

Performance improvements:

• If nrow > ncol don't transpose since the next step also transposes. Only transpose if nrow < ncol
• Binarize a sparse matrix instead of the dense matrix, assumes that all values are positive and anythign > 0 is binarized to 1.
• Transpose the result matrix before making it dense

Potential further improvements:

• don't transpose the result matrix since it needs to be transposed again as input to compartment calling

Benchmark:

mat <- assay(fte_bigwig)
dim(mat)
# [1] 3088298     305

# old function
system.time(old <- transformTFIDF(mat))
#    user  system elapsed
#  22.034   3.807  25.950

# new function
system.time(new <- transformTFIDF(mat))
#    user  system elapsed
#   7.456   1.041   8.537

identical(old, new)
# [1] TRUE