784.txt

This complete matlab for neural network

发信人: fervvac (高远), 信区: DataMining
标  题: Re: 新想法！
发信站: 南京大学小百合站 (Sun Nov 10 13:28:33 2002), 站内信件

I think the complexity lies in the exponential number of patterns:

Given N transctions from n items, there are 2^n distinct patterns (i.e., 
combination of items). One can easily define a "goodness" function that 
takes a pattern as input and returns a real value. If there is no good
opportunity of the function, the best one can do is to do an 
exhaustivie search. I am afraid the support/confidence measure is already
simple and friendly to faster computation. 

BTW, I remember ther are similar problems with Bayes network. Correct me
if I am wrong, :)

P.S. The above function could be made much more complicated, possibly by
taking as parameters statistics of the data/pattern, or attribute 
values associated with each transaction. That should be making the 
problem more interesting and more meaningful. But I am not aware of 
any publication on this direction yet, as I don't read much now, :(


【 在 minerboy (miner) 的大作中提到: 】
: 
: 有人考虑过支持度、信念度筐架的替换没有
: 关联挖掘之所以状态空间爆炸，为什么不与最初的筐架有关呢？
: han的书上也说了
: 满足支持度、信念度筐架的关联规则不一定就是有效的关联规则，反而会起到反作用
: 主要是因为规则的两端是负相关的
: 有人想过没有？


--
※ 来源:．南京大学小百合站 bbs.nju.edu.cn．[FROM: 143.89.41.4]