machine learning - Ideal classifiers in python to fit sparse high dimensional features (with hierarchical classification) -


this task:

i have set of hierarchical classes (ex. "object/architecture/building/residential building/house/farmhouse")--and i've written 2 ways of classifying:

  1. treating each class independently (using 1 model/classifier overall)

  2. using tree each node represents decision (the root represents "object/", , each level decreases generality), , specific model/classifier each node (here, consider c (usually 3) highest probabilities come out of each node, , propagate probabilities down (summing log probs) leaves), , choose highest.

    i had introduce way incentivize going further down tree (as stop @ object/architecture/building (if there corresponding training data)), , used arbitrary trial-and-error process decide how (i don't feel comfortable this).:

        if numcategories == 4:         tempscore +=1     elif numcategories ==5:         tempscore +=1.3     elif numcategories ==6:         tempscore +=1.5     elif numcategories >6:         tempscore +=2

it important note have around 290k training samples , ~150k (currently/mostly) boolean features (represented 1.0 or 0.0)--although it's highly sparse, use scipy's sparse matrices. also, there ~6500 independent classes (though many less each node in method 2)

with method 1, scikit's sgdclassifier(loss=hinge), around 75-76% accuracy, , linearsvc, around 76-77% (although it's 8-9 times slower).

however, second method (which think can/should perform better) neither of these classifiers produce true probabilities, , while i've attempted scale confidence scores produced .decision_functions(), didn't work (accuracies of 10-25%). thus, switched logisticregression(), gets me ~62-63% accuracy. also, nb based classifiers seem perform substantially less well.

ultimately, have twoish questions:

  1. is there better classifier (than scikit's logisticregression()) around implemented in python (could scikit or mlpy/nltk/orange/etc) can (i) handle sparse matrices, (ii) produce (something close to) probabilities, , (iii) work multiclass classification?
  2. is there way handle method 2 better? 2.a. specifically, there way better handle incentivizing classifier produce results further down tree?


Comments

Post a Comment

Popular posts from this blog

How to mention the localhost in android -

i trying call localhost in android application. unfortunately not possible used several methods emulator local host http://10.0.2.2/android/dbconnection.php and network ip address http://192.168.1.xx/android/dbconnection.php also http://localhost:8080/android/dbconnection.php but nothing connected localhost. purpose refer stackoverflow answers not working i facing same problem when worked on localhost. referred stackoverflow answer got solution create new emulator works fine in new emulator.
Read more

php - Calling a template part from a post -

i include iframe within post in wordpress. located inside content of post, meaning cannot in post template file, otherwise either before or after. i'd add facebook plugin within posts: <div id="socialwrap"><div class="fb-like" data-href="http://onlinearizonahomes.info" data-send="true" data-width="450" data-show-faces="true" data-font="arial"></div></div> i thinking of calling template part within post <?php get_template_part( 'social' ); ?> cannot use php within post without plugin (i'd avoid using plugin this). an alternative using javascript create html in post, seems unnecessary use of javascript. what best way add html within wordpress post? you can use shortcode. add functions.php function cudjex_fbshare( $atts, $content = null ) { global $post; $link = get_permalink($post->id); return '<div id="socialwrap...
Read more

c# - String.format() DateTime With Arabic culture -

check below code snippet, namespace testdateconvertion { class program { static void main(string[] args) { datetime testvalue = new datetime(2013, 12, 15, 15, 33, 44); cultureinfo culture = new cultureinfo("ar-sa"); string stringvalue = string.format(culture, "{0:d} {0:hh:mm:ss}", testvalue); console.writeline(stringvalue); console.readline(); } } } it gives output 22/02/35 15:33:44 i not getting how possible. whats 35 doing in output there that's 13/12/2015 in hijri, 35 means 1435 in hijri, default date format saudi arabia.
Read more