[jira] [Commented] (MAHOUT-688) High Document Frequency pruning for seq2sparse

["Vasil Vasilev (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MAHOUT-688?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13030477#comment-13030477 ] 

Vasil Vasilev commented on MAHOUT-688:
--------------------------------------

Hi Grant,

Thanks for contributing to the code. One remark from my side: In fact the standard deviation was intentionally calculated in such a way, because I wanted to "force" a zero mean. I.e. I want to calculate the standard deviation in such a way that the words with document frequency (DF) near to the zero have highest probability of getting in. I.e. I imagine that for every word DF there is a -DF (DF with the opposite sign) and calculate the standard deviation in such a way. This ensures that only high DF words will be pruned.

Regards, Vasil

> High Document Frequency pruning for seq2sparse
> ----------------------------------------------
>
>                 Key: MAHOUT-688
>                 URL: https://issues.apache.org/jira/browse/MAHOUT-688
>             Project: Mahout
>          Issue Type: Improvement
>            Reporter: Vasil Vasilev
>            Assignee: Grant Ingersoll
>            Priority: Minor
>              Labels: Vectorization
>             Fix For: 0.6
>
>         Attachments: MAHOUT-688.patch, MAHOUT-688.patch
>
>
> This improvement allows to prune the words with high document frequencies from the tf and tf-idf vectors produced by seq2sparse, based on the standard deviation of the words' document frequencies and specifying which rods to be pruned in a means of times this standard deviation. One good option is 3 times the standard deviation

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira