This function performs feature normalization according to user- specified parameters.

normalize.feat(siamcat, norm.method = c("rank.unit", "rank.std", "log.std",
  "log.unit", "clr"), norm.param = list(log.n0 = 1e-08, sd.min.q = 0.1, n.p =
  2, norm.margin = 1), verbose = 1)

Arguments

siamcat

an object of class siamcat

norm.method

string, normalization method, can be one of these: 'c("rank.unit", "rank.std", "log.std", "log.unit", "clr")

norm.param

list, specifying the parameters of the different normalization methods, see details for more information

verbose

control output: 0 for no output at all, 1 for only information about progress and success, 2 for normal level of information and 3 for full debug information, defaults to 1

Value

an object of class siamcat

Details

There are five different normalization methods available:

  • "rank.unit" converts features to ranks and normalizes each column (=sample) by the square root of the sum of ranks

  • "rank.std" converts features to ranks and applies z-score standardization

  • "clr" centered log-ratio transformation (with the addition of pseudocounts)

  • "log.std" log-transforms features (after addition of pseudocounts) and applies z-score standardization

  • "log.unit" log-transforms features (after addition of pseudocounts) and normalizes by features or samples with different norms

The list entries in "norm.param" specify the normalzation parameters, which are dependant on the normalization method of choice:

  • "rank.unit" does not require any other parameters

  • "rank.std" requires sd.min.q, quantile of the distribution of standard deviations of all features that will be added to the denominator during standardization in order to avoid underestimation of the standard deviation, defaults to 0.1

  • "clr" requires log.n0, which is the pseudocount to be added before log-transformation, defaults to NULL leading to the estimation of log.n0 from the data

  • "log.std" requires both log.n0 and sd.min.q, using the same default values

  • "log.unit" requires next to log.n0 also the parameters n.p and norm.margin. n.p specifies the vector norm to be used, can be either 1 for x/sum(x) or 2 for x/sqrt(sum(x^2)). The parameter norm.margin specifies the margin over which to normalize, similarly to the apply-syntax: Allowed values are 1 for normalization over features, 2 over samples, and 3 for normalization by the global maximum.

The function allows to perform a frozen normalization on a different dataset. After normalizing the first dataset, the output list $par contains all parameters of the normalization. Supplying this list together with a new dataset will normalize the second dataset in a comparable way to the first dataset (e.g. by using the same mean for the features for z-score standardization)