[
https://issues.apache.org/jira/browse/MAHOUT-1531?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13985537#comment-13985537
]
Andrew Palumbo commented on MAHOUT-1531:
----------------------------------------
Apologies.. There is some literature that discusses optimizing a single
smoothing parameter for all words later cited by Rennie:
http://stat.rutgers.edu/home/tzhang/papers/ir01_textcat.pdf (Sec 4.1)
Although it doesn't exactly follow the weight calculation outlined in Rennie,
there could be cases where it could be useful to allow the user to set ALPHA_I.
I'll make the necessary changes to the docs.
Please disregard this JIRA.
> Remove ALPHA_I CLI option in NB
> -------------------------------
>
> Key: MAHOUT-1531
> URL: https://issues.apache.org/jira/browse/MAHOUT-1531
> Project: Mahout
> Issue Type: Bug
> Components: Classification, CLI
> Affects Versions: 0.9
> Environment:
> Reporter: Andrew Palumbo
> Priority: Trivial
> Fix For: 1.0
>
> Attachments: MAHOUT-1531.patch
>
>
> The APLHA_I CLI option accepts a float for what should (according to the
> paper) be a vector of floats (or more correctly a path to a SequenceFile
> containing word, alpha_i pairs for each word in the respective dictionary).
> The current MapReduce NB implementation treats APLHA_I as a constant float
> where it should be a vector (ALPHA) of floats (smoothing parameter per word).
>
> However as (Rennie et al.) sec 2.2 uses a constant ALPHA_I = 1 and the NB
> implementation defaults to APLHA_I = 1 and is an implementation of Rennie,
> there is not really a problem.
> To avoid confusion, I think its a good idea to remove the option to set
> ALPHA_I and to keep ALPHA_I a constant 1.0 in NB.
>
--
This message was sent by Atlassian JIRA
(v6.2#6252)
