> > Hmmm, the reality is that I don't think a binary prediction will
...
> > the relative probabilities of success and failure?
> Could we perhaps encode all of the samples as binary {0,1} data, but
> train the classifier to return a continuous result? That is, use
> samples like:
>
> ((time=17411,key=0x3a4b,htl=12), target=1) // success
> ((time=17480,key=0x3a4b,htl=15), target=1) // success
> ((time=17485,key=0x3a4c,htl=8), target=1) // success
> ((time=17487,key=0x3a4b,htl=9), target=0) // fail
>
> with the regression mode. The predictor will then return values
> between 0 and 1 which we can interpret as a probability of success.
Yes I think that's doable. I've agree it makes sense to encode
things like "host ip" and "key" as binary inputs, so that the SVM
can get a sense of certain class A / class B / class C domains, or
the notion of "close keys." For time, I think it makes sense to
break it down into the standard minute:hour:weekday:monthday:month:year
format and use each of these as seperate input vectors, as there must
be some usage cycles highly correlated to some of these indicators.
Rudi
_______________________________________________
devl mailing list
[EMAIL PROTECTED]
http://hawk.freenetproject.org:8080/cgi-bin/mailman/listinfo/devl
