Logistic Accuracy tests
cicdw opened this issue · comments
We need to determine a robust testing framework for our logistic regression algorithms, starting with a review of the accuracy of the current ADMM code.
Unregularized Problems
- For unregularized problems with an intercept, we can start with the high-level test
y.sum() = sigmoid(X.dot(beta)).sum()
- For unregularized problems without an intercept, we can create a handful of tiny datasets to test on, or use well-studied public datasets (e.g., iris)
Regularized Problems
- As @hussainsultan already did, we can start with some high level 'marginal' tests which test extreme values for the various input parameters (e.g., when the regularization parameter is 'large' for l1-problems, the coefficients should all be 0)
- For regularized problems, we might need to hand craft a few tiny datasets to test on for accuracy
See #12
For unregularized problems, we can test for optimality of arbitrary problems easily by using y = sigmoid(X.dot(beta))
(no thresholding) and testing that the estimated coefficients are close to beta
; how do we reliably test regularized problems though? I've considered dual certificates, and possibly 0 (sub)gradients, but is there a better way?
cc: @mcg1969
Some examples of the issues we face in implementing a robust testing framework can be found here: https://github.com/dask/dask-glm/blob/master/notebooks/AccuracyBook.ipynb
Yeah, some sort of dual or subgradient criterion is likely needed...
https://gist.github.com/moody-marlin/e2de54ca17d615b263f80372031cb865 cc: @mpancia
Proximal grad does worst because the line search is currently very crude.
I need to find support for this, but in Stephen P. Boyd's (Stanford EE Dept) NIPS workshop about ADMM on Jan 25th 2012 he mentioned that ADMM without regularization is fragile. Boyd said that what makes it very robust and guaranteed to converge is regularization. This is mentioned somewhere before the 18 minute mark. I will find a source for this. http://videolectures.net/nipsworkshops2011_boyd_multipliers/
I have been using this site as a source of information on ADMM
http://web.stanford.edu/~boyd/papers/admm_distr_stats.html