10.txt

This complete matlab for neural network

发信人: fervvac (高远), 信区: DataMining
标  题: Re: Ask for your suggestion (About SVM)
发信站: 南京大学小百合站 (Fri Jun 14 23:39:32 2002), 站内信件

I have little background knowledge in classification and none in SVM.
However, intuitively, is your method reasonable?

Frist, if the classifier libsvm built is a decent classifier, it should make 
negative predication on most data in NP. Then will the absolute distance 
to THIS classification boundary meaningful?

Second, how to choose the weight assigned to N and NP? I don't know how 
these weight are used in SVM, but will the result be affected by that 
parameter?

This problem seems a hybrid of supervised and unsupervised learning. I 
wonder if there are already results for such cases?


【 在 strider (怎能没了斗志) 的大作中提到: 】
: 诸位大牛
: When doing my study, I bumped into a problem.  I would like to discribe this
: problem here, and then present  an early thought on this problem.  I hope get
: your suggestion on this issue .  Also, I hope it would not waste your much
: time.
: The following is my problem:
: ------------------------------------------------------------
: (Input)
: we have two sets of sample: one set consists of positive examples (labeled as
: "+"), here we denote the set as P;  the either set(here we denote it as NP)
: consists of BOTH Positive AND Negative examples, but we don't know the exact
: label of each example in this set(i.e. the examples in this set are all
: unlabeled.)
: (Output)
: I want to find out a samll proportion of examples(here we denote this set of
: example as N') from NP, so that I can consider WITH HIGH CONFIDENCE that the
: examples in N' are all negative.  In other words, I want to find a number of
: "strongest" negative examples from NP.
: (Procedure)
: How?
: (以下引言省略 ... ...)

--
※ 来源:．南京大学小百合站 bbs.nju.edu.cn．[FROM: 饮水思源BBS]