分子拓扑指数介绍.ppt

许禄 中国科学院长春应用化学研究所 六. Methods for variable selection ? Classical methods a) Forward selection b) Backward elimination c) Stepwise regression ? Leaps-and-bounds regression ? Orthogonal descriptors ? Genetic algorithm Since the three classical methods are known well and many papers on genetic algorithm have been published, thus, next I would want to only introduce the two methods as above mentioned: ? Leaps-and-bounds regression ? Orthogonal descriptors 1. Leaps-and-bounds regression This method is based on the fundamental inequality, RSS (A) ? RSS (Ai) RSS: residual sum of squares; A: any set of independent variables; Ai: a subset of A. e.g.: set A1 contains 3 variables; RSS=596; set A2 contains 4 variables; RSS=605. Thus, all the subsets of A2 will be ignored, because of these subsets with RSS greater than that for A2, and also for A1. 2. Orthogonal algorithm (cf. M. Randic, New J. Chem., 15(1991)517) (Gram-Schmidt 正交法) 若变量集X有N个变量按一定规则排序后为X1, X2, …XN, 第一步, 取X1为第一个正交基Ω1, 使X2, …XN和X1正交. 以X2为例, 用X1作自变量, X2为因变量进行一元回归, 得X21=X2 (实际值)– X2(计算值), 和X2一样由其它变量得到Xi1. 第二步, 取X21为正交基Ω2, 使其它变量Xi1和Ω2正交. 重复如上过程直到得到N个正交变量Ω1, Ω2,…ΩN. 由于变量相互间彼此正交, 因而, 各变量作用可以单独测试, 而不受其它变量的影响, 只要对正交变量Ω1, Ω2,…ΩN和因变量Y进行一元回归计算出回归系数Ri, 即可以计算出它们之间相互组合后的回归系数R, 公式如下: 式中Ri 为正交化的变量i与性质间的相关系数。即，首先将原变量正交化，并按与性质相关系数大小排序，然后，很易于进行最优变量子集的组合。 We have data x1, x2, x3 and x4 (the raw data are not given here in detail). The following are the results of regressions. On the up half, the regressions are performed with x1, x2, x3 and x4. On the down half, the regressions are performed with the orthogonal variables ?1 , ?2,, ?3, ?4. ---------------------------------------------------------------------------------------------------------------- Coefficients of regression equations -------