Mean imputation for linear algebra?

[kose-y]

Currently, linear algebra routines impute any missing value with zero. I know it's going to be a performance killer, but isn't imputing it with the mean more natural?

[Hua-Zhou]

If it can be done efficiently, certainly I prefer imputation by mean.

On Thu, Jul 30, 2020 at 4:40 AM Seyoon Ko notifications@github.com wrote:

Currently, missing values in linear algebra routines impute any missing value with zero. I know it's going to be a performance killer, but isn't imputing it with the mean more natural?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/OpenMendel/SnpArrays.jl/issues/67, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABGAPMJFWJ3XKCQ43VTC43TR6FL4RANCNFSM4PNRKHMA .

[kose-y]

@Hua-Zhou Quick benchmark says that mean imputation is 25-30% slower than zero imputation on CPU (still twice as fast as single-threaded BLAS) and is not slower on GPU. Should I make it a default behavior?

[Hua-Zhou]

Yes please make it default. Please also update the function documentation and package documentation.

On Thu, Jul 30, 2020 at 4:48 PM Seyoon Ko notifications@github.com wrote:

@Hua-Zhou https://github.com/Hua-Zhou Quick benchmark says that mean imputation is 25-30% slower than zero imputation on CPU (still twice as fast as single-threaded BLAS) and is not slower on GPU. Should I make it a default behavior?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/OpenMendel/SnpArrays.jl/issues/67#issuecomment-666780473, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABGAPMK6SAYBIRQSMIVLWH3R6IBGDANCNFSM4PNRKHMA .

[kose-y]

OK. I will finish it off today.