spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stephen Boesch <java...@gmail.com>
Subject Re: Multinomial regression with spark.ml version of LogisticRegression
Date Sun, 29 May 2016 22:48:45 GMT
Thanks Bryan for that pointer : I will follow it. In the meantime the One
vs Rest appears to satisfy the requirements.

2016-05-29 15:40 GMT-07:00 Bryan Cutler <cutlerb@gmail.com>:

> This is currently being worked on, planned for 2.1 I believe
> https://issues.apache.org/jira/browse/SPARK-7159
> On May 28, 2016 9:31 PM, "Stephen Boesch" <javadba@gmail.com> wrote:
>
>> Thanks Phuong But the point of my post is how to achieve without using
>>  the deprecated the mllib pacakge. The mllib package already has
>>  multinomial regression built in
>>
>> 2016-05-28 21:19 GMT-07:00 Phuong LE-HONG <phuonglh@gmail.com>:
>>
>>> Dear Stephen,
>>>
>>> Yes, you're right, LogisticGradient is in the mllib package, not ml
>>> package. I just want to say that we can build a multinomial logistic
>>> regression model from the current version of Spark.
>>>
>>> Regards,
>>>
>>> Phuong
>>>
>>>
>>>
>>> On Sun, May 29, 2016 at 12:04 AM, Stephen Boesch <javadba@gmail.com>
>>> wrote:
>>> > Hi Phuong,
>>> >    The LogisticGradient exists in the mllib but not ml package. The
>>> > LogisticRegression chooses either the breeze LBFGS - if L2 only (not
>>> elastic
>>> > net) and no regularization or the Orthant Wise Quasi Newton (OWLQN)
>>> > otherwise: it does not appear to choose GD in either scenario.
>>> >
>>> > If I have misunderstood your response please do clarify.
>>> >
>>> > thanks stephenb
>>> >
>>> > 2016-05-28 20:55 GMT-07:00 Phuong LE-HONG <phuonglh@gmail.com>:
>>> >>
>>> >> Dear Stephen,
>>> >>
>>> >> The Logistic Regression currently supports only binary regression.
>>> >> However, the LogisticGradient does support computing gradient and loss
>>> >> for a multinomial logistic regression. That is, you can train a
>>> >> multinomial logistic regression model with LogisticGradient and a
>>> >> class to solve optimization like LBFGS to get a weight vector of the
>>> >> size (numClassrd-1)*numFeatures.
>>> >>
>>> >>
>>> >> Phuong
>>> >>
>>> >>
>>> >> On Sat, May 28, 2016 at 12:25 PM, Stephen Boesch <javadba@gmail.com>
>>> >> wrote:
>>> >> > Followup: just encountered the "OneVsRest" classifier in
>>> >> > ml.classsification: I will look into using it with the binary
>>> >> > LogisticRegression as the provided classifier.
>>> >> >
>>> >> > 2016-05-28 9:06 GMT-07:00 Stephen Boesch <javadba@gmail.com>:
>>> >> >>
>>> >> >>
>>> >> >> Presently only the mllib version has the one-vs-all approach
for
>>> >> >> multinomial support.  The ml version with ElasticNet support
only
>>> >> >> allows
>>> >> >> binary regression.
>>> >> >>
>>> >> >> With feature parity of ml vs mllib having been stated as an
>>> objective
>>> >> >> for
>>> >> >> 2.0.0 -  is there a projected availability of the  multinomial
>>> >> >> regression in
>>> >> >> the ml package?
>>> >> >>
>>> >> >>
>>> >> >>
>>> >> >>
>>> >> >> `
>>> >> >
>>> >> >
>>> >
>>> >
>>>
>>
>>

Mime
View raw message