machine learning - Scikit-learn: BernoulliNB, v0.10 vs v0.13: very different results -

-

this of follow-up this thread, getting erroneous results gaussiannb classifier, turned out because had scikit-learn v0.10 on linux vm doing experiments on. ended using bernoulli , multinomial nb classifiers instead, when (finally) got scipy installed on macbook, scikit-learn version grabbed 0.13, latest of writing. presented new problem:

  • on v0.10, getting on 90% accuracy bernoullinb classifier on 1 of feature sets, notable improvement i've gotten far.
  • on v0.13, it's coming in @ 67% using same code

does know changed between versions? had @ repo history didn't see account kind of change in accuracy. since i'm getting results bernoullinb v0.10, i'd use them, i'm hesitant without little more understanding of conflicting results between versions.

i've tried setting (newer) class_prior property didn't change results 0.13.

edit: short of coming worked example (which will, well, work on), 0.13 outcomes heavily biased, not expect bayesian classifier, , leads me believe may have been regression on class prior calculations, though haven't tracked down yet. example:

0.10: t\p f   m f   120 18 m   19  175  0.13: t\p f   m f   119 19 m   59  135

edit 2:

i worked through few examples hand. 0.13 version correct , 0.10 version not, both suspected , feared. error in 0.10 appears in class prior calculation. _count function bugged, on this line of file, class counts wrong: compare the 0.13 branch, ignoring 2 branches pull in smoothing factors @ different places.

i have think more, why botched feature counts resulting in such performance on data, , i'm still little unsure why setting class priors didn't work. perhaps penalizing against male bias present in source documents?

edit 3:

i believe doing. _count function, , consequently calculation of feature priors within fit, not take parameter effect, while class_priors taken account within predict, not used build model during training. not sure if intentional - want ignore priors used build model @ testing time?

to sum results, bug in 0.10 version of bernoullinb classifier, skewing class counts when calculating feature priors, , apparently biasing resulting model yield superior results. managed adapt pieces of doing , got equivalent performance (correct) multinomialnb in version 0.13.


Comments

Post a Comment

Popular posts from this blog

Perl - how to grep a block of text from a file -

-
it can in xml or text format. how in general grep block of text in perl? <track type="ws"> <range> <rangestart>0</rangestart> <rangeend>146.912</rangeend> <locationindex>0</locationindex> <propertyindex>0</propertyindex> </range> </track> <track type="ps" id="1"> <range> <rangestart>0</rangestart> <rangeend>146.912</rangeend> <locationindex>1</locationindex> <propertyindex>1</propertyindex> </range> </track> i want grep type="ps" , till </range> . one solution open file, read line line , match block. open(fh, "file.txt"); foreach $line (<fh>) { if ($line =~ m/type="cc"(.*?)<\/range>/) { print $1; } } but there more optimal solution without reading file line l
Read more

delphi - How to remove all the grips on a coolbar if I have several coolbands? -

-
Image
if set mycoolbar.fixedorder true,only grip on first band hidden. well,if use delphi 7 create vcl forms application,then put coolbar on , create 3 coolbans hold other controls,only grip , top coolbands can hidden setting mycoolbar.fixedorder:=true. i've uploaded picture make things clear. you got fixedorder property wrong. fixed property not allow bands rearranged if set true . setting property of coolbar true keep user changing bands order @ runtime, user can still move , resize bands. can give advice can do, actual solution problem, well, have wait answer. my advice use 3 coolbars in row , setting "fixedorder" property true , bandborderstyle bsnone . way grip hidden on of them. about property, it's not bug of ide, actual preference of property.
Read more

javascript - Animating array of divs; only the final element is modified -

-
this baffling me @ moment. i'm experienced programmer looking throw bit of surprise fiancee on countdown page our wedding. desired effect simulated confetti falling behind main page. setup follows: i have array of divs in javascript. each div absolutely positioned on page. have interval set updates every 50 or milliseconds. on each tick, run loop on array of divs, , calculations , update positions. problem problem having, however, though can see divs created, stored array, , modified (in case, each has slight randomized rotation) during initialization phase, once tick update position starts, div stored in whatever last slot of array gets position updated. tried hard-coding each div index testing purposes, strictly last element update position. upon printing current "left" , "top" style values each div before , after calculations shows value has been changed, no visible change noticeable. here's basic code: javascript var container;
Read more