biglasso
biglasso copied to clipboard
Will this package supports a weight vector?
Impressive package, would love to be able to see support for weights
@jeffwong-nflx , thanks for your comments. Currently, "screen = SSR" supports weights options. "screen = SSR-BEDPP" for weight support is under development. (So I think I will need to disable weight option for other screening rules, will fix this!)
I was just wondering if there is any update planned to also support observation weights? Is it the case that right now, for family="gaussian", one could pre-multiply X by diag(sqrt(weights)) and y by sqrt(weights) to support regularised weighted OLS? Only thing is that cv.biglasso would not be correct then, as the evaluation metric would also have to be the weighted mean squared error instead of the unweighted mean squared error... Also for the binomial case, observation weights would be really useful. I understand that for binomial models, biglasso now just allows for an outcome variably y that is a vector of 0 or 1s. But in many cases, one would be dealing with aggregate counts for the two categories, which could be supported by allowing outcome variable y to represent the observed proportion and the observation weights to be the total count over both categories...