spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Franco Barrientos" <franco.barrien...@exalitica.com>
Subject RE: Effects problems in logistic regression
Date Mon, 22 Dec 2014 13:27:28 GMT
Thanks again DB Tsai, LogisticRegressionWithLBFGS works for me!

 

De: Franco Barrientos [mailto:franco.barrientos@exalitica.com] 
Enviado el: jueves, 18 de diciembre de 2014 16:42
Para: 'DB Tsai'
CC: 'Sean Owen'; user@spark.apache.org
Asunto: RE: Effects problems in logistic regression

 

Thanks I will try.

 

De: DB Tsai [mailto:dbtsai@dbtsai.com] 
Enviado el: jueves, 18 de diciembre de 2014 16:24
Para: Franco Barrientos
CC: Sean Owen; user@spark.apache.org <mailto:user@spark.apache.org> 
Asunto: Re: Effects problems in logistic regression

 

Can you try LogisticRegressionWithLBFGS? I verified that this will be converged to the same
result trained by R's glmnet package without regularization. The problem of LogisticRegressionWithSGD
is it's very slow in term of converging, and lots of time, it's very sensitive to stepsize
which can lead to wrong answer. 

 

The regularization logic in MLLib is not entirely correct, and it will penalize the intercept.
In general, with really high regularization, all the coefficients will be zeros except the
intercept. In logistic regression, the non-zero intercept can be understood as the prior-probability
of each class, and in linear regression, this will be the mean of response. I'll have a PR
to fix this issue.





Sincerely,

DB Tsai
-------------------------------------------------------
My Blog: https://www.dbtsai.com
LinkedIn: https://www.linkedin.com/in/dbtsai

 

On Thu, Dec 18, 2014 at 10:50 AM, Franco Barrientos <franco.barrientos@exalitica.com <mailto:franco.barrientos@exalitica.com>
> wrote:

Yes, without the “amounts” variables the results are similiar. When I put other variables
its fine.

 

De: Sean Owen [mailto:sowen@cloudera.com <mailto:sowen@cloudera.com> ] 
Enviado el: jueves, 18 de diciembre de 2014 14:22
Para: Franco Barrientos
CC: user@spark.apache.org <mailto:user@spark.apache.org> 
Asunto: Re: Effects problems in logistic regression

 

Are you sure this is an apples-to-apples comparison? for example does your SAS process normalize
or otherwise transform the data first? 

 

Is the optimization configured similarly in both cases -- same regularization, etc.?

 

Are you sure you are pulling out the intercept correctly? It is a separate value from the
logistic regression model in Spark.

 

On Thu, Dec 18, 2014 at 4:34 PM, Franco Barrientos <franco.barrientos@exalitica.com>
wrote:

Hi all!,

 

I have a problem with LogisticRegressionWithSGD, when I train a data set with one variable
(wich is a amount of an item) and intercept, I get weights of

(-0.4021,-207.1749) for both features, respectively. This don´t make sense to me because
I run a logistic regression for the same data in SAS and I get these weights (-2.6604,0.000245).

 

The rank of this variable is from 0 to 59102 with a mean of 1158.

 

The problem is when I want to calculate the probabilities for each user from data set, this
probability is near to zero or zero in much cases, because when spark calculates exp(-1*(-0.4021+(-207.1749)*amount))
this is a big number, in fact infinity for spark.

 

How can I treat this variable? or why this happened? 

 

Thanks ,

 

Franco Barrientos
Data Scientist

Málaga #115, Of. 1003, Las Condes.
Santiago, Chile.
(+562)-29699649 <tel:%28%2B562%29-29699649> 
(+569)-76347893 <tel:%28%2B569%29-76347893> 

franco.barrientos@exalitica.com <mailto:franco.barrientos@exalitica.com>  

www.exalitica.com <http://www.exalitica.com/> 


  <http://exalitica.com/web/img/frim.png> 

 


Mime
View raw message