[R] Glmnet Logistic Variable Questions

[Ben Bolker]

On 11-10-25 01:35 PM, julien giami wrote:
> The reason i use glmnet is that it makes the handling of 400,000
> observations easier to handle in terms of memory,
> 
> I am looking on sparse matrices but i dont understand how to build
> interacting using sparse matrices
> 

  If you're not familiar with glmnet but you are familiar with GLMs in
general may I suggest bigglm() in the biglm package?
> 
> 
> On Tue, Oct 25, 2011 at 12:34 PM, Marc Schwartz <marc_schwartz at me.com> wrote:
>>
>> On Oct 25, 2011, at 11:16 AM, Ben Bolker wrote:
>>
>>> Bert Gunter <gunter.berton <at> gene.com> writes:
>>>
>>>>
>>>> If I understand you correctly, it sounds like you need to do some reading.
>>>>
>>>> ?lm and ?formula tell you how to specify linear models for glm or glmnet.
>>>> However, if you do not have sufficient statistical background, It probably
>>>> will be incomprehensible, in which case you should consult your local
>>>> statistician.
>>>>
>>>> For glmnet, go to the linked references given in the Help file.There is no
>>>> such thing as AIC for these models, as they are penalized fits (with users
>>>> choosing the penalization tradeoff). Again, consult your local statistician
>>>
>>>  Let me second Bert's concern, but in the meantime, if what you
>>> want are *all two-way interactions among variables, you can follow
>>> this example:
>>>
>>>> d <- data.frame(y=runif(100),x1=runif(100),x2=runif(100),x3=runif(100))
>>>> gg <- lm(y~(.)^2,data=d)
>>>> names(coef(gg))
>>> [1] "(Intercept)" "x1"          "x2"          "x3"          "x1:x2"
>>> [6] "x1:x3"       "x2:x3"
>>>
>>>
>>> I have done the example with continuous variables and with lm() here,
>>> but it should generalize easily to (1) a mixture of categorical and
>>> continuous variables and (2) other R modeling functions
>>
>>
>>
>> There is a difference with glmnet however vis-à-vis its handling of factors. There is a recent discussion here:
>>
>>  https://stat.ethz.ch/pipermail/r-help/2011-August/285905.html
>>
>> which covers the topic. Be sure to read the replies, including Martin's.
>>
>> HTH,
>>
>> Marc Schwartz
>>
>>