« first day (2173 days earlier)   

vzn
1:32 AM
hi anyone with experience with ruby statsample? a big hassle so far, lots of gory details below :(
 
2:20 AM
@vzn All you real need to solve regression is a linear equation solver.
This post seems to point at one: flylib.com/books/en/2.44.1.39/1
Note though, that in that book they take the inverse of the coefficient matrix and multiply by the right hand side.
That's mathematically correct, but you shouldnt do it that way computationally.
Instead of X^{-1} y do solve(X, y).
 
vzn
3:00 AM
@MatthewDrury yeah thx. it seems to be using Matrix include which is presumably quite similar to one that comes with statsample and presumably also a "gem" which requires ruby install... afaik linear eqn solvers are also gem libraries in ruby... (maybe even the same ones used by statsample...) actually am more looking for a ML like library in ruby that include regression, not sure if one exists...
does anyone around have experience with python statistics/ ML? wondering how it compares...
 
Python is pretty great. It has a strong scientific computing community.
sklearn and statsmodels will meet many needs. For bigger datasets, you can use something like H2O from python pretty easily.
 
 
7 hours later…
9:48 AM
What does this learning curve tell you guys beyond massive overfitting? imgur.com/a/gQZRw
 
 
2 hours later…
11:20 AM
@IuliusCurt I don't think you can see overfitting from that chart. You would need to plot model fit by model complexity for a fixed set of training data to see that.
Just having a difference between training data and hold out data fit is not in itself indicative of an overfit model.
All I see in that plot is that you could stand to have even more than ~1000 training examples for whatever fixed model you are fitting.
The fact that training "score" is 1 for every size of training data is probably interesting, but I don't know what "score" is, so it's hard to interpret.
 
 
1 hour later…
12:28 PM
@MatthewDrury "Score" is mean accuracy (Jaccard)
Basically, not a single training sample is mistaken (no outlier there? hard to believe). That's what makes me think of overfitting
 

« first day (2173 days earlier)