mahout-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ted Dunning <ted.dunn...@gmail.com>
Subject Re: linear regression package in mahout
Date Mon, 12 Apr 2010 03:21:13 GMT
Never compute the inverse of a matrix.  Use QR or SVD decompositions for
least squared error problems or an optimization technique for convex
problems.

What you have is a small under-determined system that can be easily handled
using a package like R.  You don't need to worry about scaling.  You need to
worry (A LOT) about over-fitting.

Normal OLS will fail disastrously on the problem as you state it.  It might
be possible to use a regularization technique, but with only 10 data points
and 200 parameters to fit, you are unlikely to succeed unless you know a LOT
about your problem that you can encode as a prior.



On Sun, Apr 11, 2010 at 5:32 PM, prasenjit mukherjee
<prasen.bea@gmail.com>wrote:

> I am trying to compute regression coefficients, where the dimensions
> are ~ 200 and number of points = 10. Basically I need to compute the
> beta matrix  using OLS ( refer
> http://en.wikipedia.org/wiki/Ordinary_least_squares ). The main
> bottleneck seems to be computing the inverse of a 200X200 matrix.
>
> Any pointers/suggestions ?
>
> -Thanks,
> Prasen
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message