[SciPy-user] 'linear regression'

[Alan G Isaac]

On Sun, 05 Jun 2005, David Grant apparently wrote: 
> http://www.davidgrant.ca/python/bestFitLine.py 

1. For something like this I think GPL is major overkill. 
   How about public domain, or at least MIT? (The comments 
   below are in the public domain.  ;-)) 
2. In the loop 
        for i in range(len(x)): 
                data[i] = (x[i],y[i],(max(abs(y))+min(abs(y)))/100.) 
   I think you could: 
   - omit the weight. In any case, you chose a weight that 
     does not look generally optimal (but will not usually 
     hurt or help). 
   - compute the weight outside the loop 
   - let data=zip(x,y,repeat(wt)) (for whatever weight you 
     choose, if you use one) (repeat is in itertools)
3. For your stated purpose, if I understood, you do not need 
   nonlinear LS and could use a linear solution (noniterative). 
   (Scipy has this.) 
4. An object oriented approach seems more natural here, 
   perhaps, especially given your reliance on Scientific 
   Python. 

Point 2 seems to raise a question for Konrad: might allowing 
leastSquaresFit to accept an iterator instead of a list as 
data have some advantages? 

Cheers, 
Alan Isaac