ffx
ffx copied to clipboard
natekupp
Fails to identify y=x^2
Added a test [5510110672e86ddd5620b432574af6e0c7136a8a] with data y=x^2 for x=[0, 1, 2, 3]. FFX does pretty well, but I expected the exact relationship. In the 4-base model, there are two x^2 terms -- could this be related to the handling of second-order bases mentioned in #5 ?
Num bases,Test error (%),Model 0, 62.4453, 3.50 1, 11.4284, 0.640 + 0.817_x^2 2, 1.6635, 0.0846 + 0.972_x^2 + 0.00984_x 4, 0.7507, (0.0973 + 0.523_x^2 + 0.440_x^2) / (1.0 - 0.00214_x - 0.00168*x)
Maybe this is not the fault of FFX. The elastic net just doesn't seem to do well modelling this type of data (no noise, simple input-output relationship). Here is a test:
http://stackoverflow.com/questions/22738879/should-elastic-net-regression-be-able-to-regress-y-x-perfectly
The two x^2 terms seem to be coming from passing the base x into a model, which allocates it to both numerator and denominator. When it comes back, x is collected twice. It goes:
FFXModelFactory.build -> _basesToModels -> _pathwiseLearn -> _allocateToNumerDenom
The failure to find the exact model is expected, but the multiple occurrences of x^2 seems to be a flaw, so I'll leave this issue open but rename it.