Comments on: Semi-supervised Naive Bayes in Python

By: Purvi

Purvi — Sat, 26 Nov 2011 16:27:25 +0000

Hello rick wei,

As you suggest, I have provided input data for td and delta in train function, but I am getting error of objects are not aligned. Here is output of that

File “/var/test/seminb.py”, line 156, in train
self.p_w_c[w,c] += td[w,d] * delta[d,c]
File “/usr/lib/python2.6/dist-packages/numpy/core/defmatrix.py”, line 290, in __mul__
return N.dot(self, asmatrix(other))
ValueError: objects are not aligned

I am running it on ubuntu linux system.

Please advise what should i do to successfully run train and other functions.

By: Purvi

Purvi — Thu, 24 Nov 2011 10:26:56 +0000

Hello..
thanks for your code..
Can anyone give complete working code with input matrix ? Thanks in advance..

By: hoh

hoh — Thu, 10 Nov 2011 15:56:10 +0000

Does weka support Semi-Supervised Learning?

By: Mathieu

Mathieu — Thu, 22 Sep 2011 07:42:56 +0000

Just quickly looking at the equations in this post, there doesn’t seem to be any reason that it shouldn’t work. So I guess it’s probably a bug in your implementation. Try to keep in mind the big picture: first train your classifier on the labeled data, then use your current classifier to find probabilistic labels of the unlabeled data, retrain with all the data (labeled and probabilistically labeled) and repeat. Also keep your implementation simple until you get it right and optimize performance afterwards.

By: Abhaya

Abhaya — Tue, 20 Sep 2011 11:18:37 +0000

Hi Mathieu,

I am trying to use the Kamal Nigam’s algorithm for the Multi-Variate Burnoulli NB. I am running into the strange problem of likelihood of data decreasing with every EM step! I believe that I might have made a mistake in adopting the algorithm for the Burnoulli case. Do you have any thoughts on how that can be done correctly?

Regards,
Abhaya

By: Quora

Quora — Thu, 05 May 2011 16:38:58 +0000

How big the training set should be in the Naive Bayes text classification?…

* First make sure that data is balanced. Since simple naive Bayesian algorithm won’t work for unbalanced dataset. If dataset is unbalanced, then I suggest you to try out complement Bayesian algorithm. Weka(http://www.cs.waikato.ac.nz/ml/weka/) and m…

By: Mathieu

Mathieu — Fri, 22 Apr 2011 09:09:56 +0000

I think it’s a bug in your program. It means that you’re doing something like this: arr[i, j] = [1, 2, 3, ...]. You cannot set the element of an array with a sequence.

By: Rick wei

Rick wei — Thu, 21 Apr 2011 21:14:14 +0000

Hi,
Thanks for your program share.

I met an error where the number of row of numpy.array more than 1600.
“”" ValueError: setting an array element with a sequence. “”"

Does the numpy limit the size of array? If the anwer is yes, do you know how to expend the array size?

By: rick wei

rick wei — Fri, 15 Apr 2011 14:08:32 +0000

Hi,
Very appreciate for your response. It’s very helpful. Thank you.

By: Mathieu

Mathieu — Wed, 13 Apr 2011 07:11:19 +0000

In the E-step, I use the current model to label unlabeled data, i.e., compute the fractional labels for the unlabeled data.
In the M-step, I compute the counts for unlabeled data, I add the counts for labeled data (those don’t change so they are computed outside of the loop), and finally I normalize to get probabilities.

You can find a more recent version of my code here:
https://github.com/mblondel/scikit-learn/blob/semisupervised/scikits/learn/naive_bayes.py

You can find unit tests here:
https://github.com/mblondel/scikit-learn/blob/semisupervised/scikits/learn/tests/test_naive_bayes_semi.py

(Sorry I don’t have time to provide support about these programs)