wrong input into entropy function

[minhsphuc12]

scikit-learn/sklearn/metrics/cluster/supervised.py

Line 751 in 74ae6a0

h_true, h_pred = entropy(labels_true), entropy(labels_pred)

entropy must be inputed with frequency table of labels_true / labels_pred, not those array directly.
Example solution

from collections import Counter
freq_table = [x[1] for x in dict(Counter(labels_true)).items()]
entropy(freq_table)

[vasujogani]

working on it

[jnothman]

Are you sure? Our entropy implementation normalises the distribution.

[vasujogani]

I think you might be right. I just saw entropy using np.unique which I believe would take care of the issue

[minhsphuc12]

Please consider example
vec_a = np.asarray([0, 1, 2, 0, 3, 4, 5, 1])
stats.entropy(vec_a) #1.6304325826288886

from collections import Counter
freq_table = [x[1] for x in dict(Counter(vec_a)).items()]
stats.entropy(freq_table) #1.7328679513998633

As you can see, vec_a is cluster vector, does not has same length as freq_table.

The issue here is not normalizing the input (same length), but rather using cluster vector, instead of using frequency table as input.

[jnothman]

We don't use stats.entropy. our entropy does the right thing for this input.

[minhsphuc12]

Thanks. Last question can you point to the source of used entropy function here. I can not find it.

On Mon, Jul 29, 2019 at 15:31 Joel Nothman ***@***.***> wrote: We don't use stats.entropy. our entropy does the right thing for this input. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#14497?email_source=notifications&email_token=ABX74MTU7TJD7F6YFHRJFCDQB2TF7A5CNFSM4IHNFGH2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD277RRY#issuecomment-515897543>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABX74MWO5SQSHWNPGILKVTTQB2TF7ANCNFSM4IHNFGHQ> .

-- Sincerely NGUYEN Minh Phuc (Mr) Tel: (84) 1 696 111 003 Skype: nasdap1112 Facebook: http://www.facebook.com/minhsphuc12

[jnothman]

It's defined in the same file

[minhsphuc12]

Thanks problem solved. Root cause is I do not aware that averaging method used in my sklearn version is ‘max’, instead of ‘arithmetic’ of 0.22.

[minhsphuc12]

Although we would need to update the documentation example.

[thomasjpfan]

How would you want to update the example?

[cmarmo]

Closing this one as it is quite old, the main issue has been solved and there was no answer to the last question.
Feel free to reopen if things can be clarified.