SVD based method for predicting missing values in future experiments

Question

I would like to pose my question in terms of a hypothetical example, where I have a matrix:

11 12 13...1m 
21 22 23 ..2m
.  .  .    3m
.  .  .  .  .
11 n2 n3 .  nm

The matrix has no missing values. Each column corresponds to one experiment and each row corresponds to a particular measurement parameter ( eg: temperature @ point 1, temperature @ point 2... Pressure....)

next, I conduct a New experiment, and I measure most parameters but not all n parameters. I want to estimate the missing parameter(s) in my New experiment.

Is there a formula based on SVD/PCA to calculate the missing value in my New experiment based on my 'training' matrix?

Are there any better machine learning / statistical methods to perform this calculation ?

Thanks in advance!

Answer 1

See this paper: Hastie, Mazumder, Lee and Zadeh, Matrix completion and low-rank SVD via fast alternating least squares.