你知道SAS也可以實現(xiàn)神經(jīng)網(wǎng)絡(luò)嗎

神經(jīng)網(wǎng)絡(luò)的理論太長了，我就不寫上來了，本次的代碼是根據(jù)這本書《數(shù)據(jù)挖掘與應(yīng)用》--張俊妮中神經(jīng)網(wǎng)絡(luò)這一章我做了思路的改動以及在原本代碼的基礎(chǔ)上，我把它封裝好變成一個完整的宏。如果后面你們需要去了解sas神經(jīng)網(wǎng)絡(luò)這個proc neural過程也可以去購買這本來讀，這本書沒有配套的代碼，所以代碼也是我一個一個照著書敲來之后更改的。

說下這個代碼的思路。

1、宏的第一步是采用神經(jīng)網(wǎng)絡(luò)建立廣義線性模型，沒有隱藏層，對數(shù)據(jù)做第一次的訓(xùn)練。這里的數(shù)據(jù)宏里面已經(jīng)自動拆分成測試和訓(xùn)練了，你自己不用拆了。

2、算出神經(jīng)網(wǎng)絡(luò)建立廣義線性模型的結(jié)果算出原始數(shù)據(jù)（訓(xùn)練集以及測試集）中每一個，客戶的違約概率，之后算其ks值。

3、接下來就是循環(huán)隱藏層，從1循環(huán)到3，當(dāng)然你要是覺得3層太少，你可以再設(shè)置，使用的是早停止（張俊妮的《數(shù)據(jù)挖掘與應(yīng)用》的114頁種有詳細(xì)解釋這個算法）法建立多層感知模型，那么這里第一次當(dāng)然循環(huán)就是1層啦，那就是1層感知模型。

4、在當(dāng)隱藏層為一層的時候，我們會擬合兩次神經(jīng)網(wǎng)絡(luò)，第一次不輸出結(jié)果，只是產(chǎn)出在隱藏層為一層的時候，挑選出最優(yōu)的變量權(quán)重，擬合一個使用早停止法擬合出來的一個隱藏層為一層的神經(jīng)網(wǎng)絡(luò)模型，利用出來的變量規(guī)則，算出客戶的概率之后算出模型的ks值。

5、到這里并不是要循環(huán)隱藏層為兩層，還有呢，別著急，這時候隱藏層為一層的前提下，再使用，規(guī)則化法建模一層感知器模型，剛才的一層隱藏層使用的早停止法，現(xiàn)在使用的是規(guī)則化法，這時候規(guī)則化法去的權(quán)衰減常數(shù)的四種取值（規(guī)則化法也可以在書里的115頁看到。）四種取值是：0.1 、0.01、0.001、0.0001然后循環(huán)之后算出，每個模型的ks記錄。

6、所以循環(huán)一次隱藏層的層數(shù)，是得到4個模型的，早停止法一個，規(guī)則化法四個。

7、再一次循環(huán)隱藏層的2、3層。最終你可以在ks的匯總跑那個表中，選出訓(xùn)練數(shù)據(jù)以及測試數(shù)據(jù)ks都高的模型，作為你最終的模型。

%macromlps(dir,data,list_varname,y_var);

proc datasets lib=work;

delete alltrainfit allvalidfit vaild_ks_total train_ks_total;

run;

data M_CALL_DAY_TOTAL4_t;

set &data.;

indic=_n_;

run;

Proc sort data=M_CALL_DAY_TOTAL4_t; by &y_var.;run;

proc surveyselect data =M_CALL_DAY_TOTAL4_t method = srs rate=0.8

out = traindata;

strata &y_var.;

run;

proc sql;

create table validdata as

select * from

M_CALL_DAY_TOTAL4_t where indic not in (select indic from traindata);

quit;

data traindata;

set traindata;

drop SelectionProb SamplingWeight indic;

run;

data validdata;

set validdata;

drop indic;

run;

proc dmdb data=traindata dmdbcat=dmcdata;

class &y_var.;

var &list_varname.;

run;

proc dmdb data=validdata dmdbcat=dmcdata;

class &y_var.;

var &list_varname.;

run;

data decisionmatrix;

&y_var.=1;

to_1=0;

to_2=1;

output;

&y_var.=0;

to_1=1;

to_2=0;

output;

run;

proc neural data=traindata validdata=validdata dmdbcat=dmcdata ranscale=0.1random=0;

input &list_varname./level=int;

target &y_var./level=nom;

decision decdata=decisionmatrix(type=loss) decvars=TO_1 TO_2;

archi glim;

nloptions maxiter=300;

train ;

code file="&dir.nncode_germancredit_glim.sas";

score data=traindata nodmdb out=traindata_GLIM outfit=trainfit_GLIM role=TRAIN;

score data=validdata nodmdb out=validdata_GLIM outfit=validfit_GLIM role=valid;

run;

data test_train(keep=appl_id &y_var.point);

set traindata;

%include"&dir.nncode_germancredit_glim.sas";

rename P_&y_var.0=point;

run;

proc npar1way data=test_train noprint;

class &y_var.;

var point;

output out=ks_t(keep=_d_ p_ksa rename=(_d_=KS p_ksa=P_value));

run;

data test_train_ks;

set ks_t;

length model$50.;

model="glim";

run;

proc append base=train_ks_total data=test_train_ks force;

run;

data vaild_train(keep=appl_id &y_var.point);

set validdata;

%include"&dir.nncode_germancredit_glim.sas";

rename P_&y_var.0=point;

run;

proc npar1way data=vaild_train noprint;

class &y_var.;

var point;

output out=ks_v(keep=_d_ p_ksa rename=(_d_=KS p_ksa=P_value));

run;

data test_vaild_ks;

set ks_v;

length model$50.;

model="glim";

run;

proc append base=vaild_ks_total data=test_vaild_ks force;

run;

%letnhidden=1;

%do%until(&nhidden.>3);

proc neural data=traindata validdata=validdata dmdbcat=dmcdata graph;

input &list_varname./level=int;

target &y_var./level=nom;

decision decdata=decisionmatrix(type=loss) decvars=TO_1 TO_2;

archi MLP hidden=&nhidden.;

nloptions maxiter=300;

train estiter=1outest=weights_MLP&nhidden._ES outfit=assessment_MLP&&nhidden._ES;

/*code file="&dir.nncode_germancredit_glim.sas";*/

/*score data=traindata nodmdb out=traindata_GLIM outfit=trainfit_GLIM role=TRAIN;*/

/*score data=validdata nodmdb out=validdata_GLIM outfit=validfit_GLIM role=valid;*/

run;

proc sort data=assessment_MLP&&nhidden._ES;

by _VALOSS_;

RUN;

DATA BESTITER;

SET assessment_MLP&&nhidden._ES;

IF _N_=1;

RUN;

proc sql;

select _iter_ into:BESTITER from BESTITER;

quit;

data bestweights;

set weights_MLP&nhidden._ES;

if _type_="PARMS"AND _iter_=&bestiter.;

drop _tech_ _type_ _name_ _decay_ _seed_ _nobj_ _obj_ _objerr_

_averr_ _vnobj_ _vobj_ _vobjerr_ _vaverr_ _p_num_ _iter_;

run;

proc neural data=traindata validdata=validdata dmdbcat=dmcdata graph;

input &list_varname./level=int;

target &y_var./level=nom;

decision decdata=decisionmatrix(type=loss) decvars=TO_1 TO_2;

archi MLP hidden=&nhidden.;

initial inest=bestweights;

train tech=none;

code file="&dir.nncode_germancredit_MLP&nhidden._ES.sas";

score data=traindata nodmdb out=traindata_MLP&Nhidden._ES outfit=trainfit_MLP&Nhidden._ES role=TRAIN;

score data=validdata nodmdb out=validdata_MLP&Nhidden._ES outfit=validfit_MLP&Nhidden._ES role=valid;

run;

data test_train(keep=appl_id &y_var.point);

set traindata;

%include"&dir.nncode_germancredit_MLP&nhidden._ES.sas";

rename P_&y_var.0=point;

run;

proc npar1way data=test_train noprint;

class &y_var.;

var point;

output out=ks_t(keep=_d_ p_ksa rename=(_d_=KS p_ksa=P_value));

run;

data test_train_ks;

set ks_t;

length model$50.;

model="ES";

run;

proc append base=train_ks_total data=test_train_ks force;

run;

data vaild_train(keep=appl_id &y_var.point);

set validdata;

%include"&dir.nncode_germancredit_MLP&nhidden._ES.sas";

rename P_&y_var.0=point;

run;

proc npar1way data=vaild_train noprint;

class &y_var.;

var point;

output out=ks_v(keep=_d_ p_ksa rename=(_d_=KS

p_ksa=P_value));

run;

data test_vaild_ks;

set ks_v;

length model$50.;

model="ES";

run;

proc append base=vaild_ks_total data=test_vaild_ks force;

run;

%letidecay=1;

%do%until(&idecay.>4);

%if&idecay.=1%then%letcedcay=0.1;

%else%if&idecay.=2%then%letcedcay=0.01;

%else%if&idecay.=3%then%letcedcay=0.001;

%else%if&idecay.=4%then%letcedcay=0.0001;

%put&cedcay.;

proc neural data=traindata validdata=validdata dmdbcat=dmcdata graph;

input &list_varname./level=int;

target &y_var./level=nom;

decision decdata=decisionmatrix(type=loss) decvars=TO_1 TO_2;

archi MLP hidden=&nhidden.;

netoptions decay=&cedcay.;

nloptions maxiter=300;

prelim5maxiter=10;

train ;

code file="&dir.nncode_germancredit_MLP&nhidden._WD&idecay..sas";

score data=traindata nodmdb out=traindata_MLP&Nhidden._WD&idecay.outfit=trainfit_MLP&Nhidden._WD&idecay.role=TRAIN;

score data=validdata nodmdb out=validdata__MLP&Nhidden._WD&idecay.outfit=validfit_MLP&Nhidden._WD&idecay.role=valid;

run;

data test_train(keep=appl_id &y_var.point);

set traindata;

%include"&dir.nncode_germancredit_MLP&nhidden._WD&idecay..sas";

rename P_&y_var.0=point;

run;

proc npar1way data=test_train noprint;

class &y_var.;

var point;

output out=ks_t(keep=_d_ p_ksa rename=(_d_=KS p_ksa=P_value));

run;

data test_train_ks;

set ks_t;

length model$50.;

model="&nhidden._WD&idecay.";

run;

proc append base=train_ks_total data=test_train_ks force;

run;

data vaild_train(keep=appl_id &y_var.point);

set validdata;

%include"&dir.nncode_germancredit_MLP&nhidden._WD&idecay..sas";

rename P_&y_var.0=point;

run;

proc npar1way data=vaild_train noprint;

class &y_var.;

var point;

output out=ks_v(keep=_d_ p_ksa rename=(_d_=KS

p_ksa=P_value));

run;

data test_vaild_ks;

set ks_v;

length model$50.;

model="&nhidden._WD&idecay.";

run;

proc append base=vaild_ks_total data=test_vaild_ks force;

run;

%letidecay=%eval(&idecay.+1);

%end;

%letnhidden=%eval(&nhidden.+1);

%end;

%mend;

/*%mlps();*/

%letlist_varname=%str(N_M5_T09_CONRT N_M6_T09_CINRT N_N5_T82_COC_RC N_M5_T03_CONRT N_M3_T10_CONRC N_M3_T83_COC_RC

N_M3_T09_CINRTN_M6_T83_COT_RC N_M3_T83_CIT_RC N_M2_T10_CONRC N_M5_T03_CINRM N_M5_T10_CONRC N_M4_T02_CONRC N_M1_T08_CONRM

N_M3_T06_CINRM N_M2_T09_CONRT N_M6_T03_CONRT N_M5_T07_CINR );

%mlps(dir=F:data_1,data=raw.CALL_HOUR2_total7_woe,list_varname=&list_varname.,y_var=y);

最終的宏里面的list_vaname就不用填了，讓他引用上面的宏list_vaname就可以了，list_vaname填的是你要去建立神經(jīng)網(wǎng)絡(luò)的變量，這里提醒一句哈，就是我嘗試了不分組，分20組，分10組，分5組的效果，我建議是將變量分組好之后再丟進去比較好，但是我說不準(zhǔn)到底是幾組好，畢竟我和你的數(shù)據(jù)不一樣。

data填的原始數(shù)據(jù)集。dir，填一個路徑，這個路徑存放的是最終的模型輸出的規(guī)則，跟決策樹那個score一個道理的。y_var填的是你的因變量。

最后看下你們最終要看的結(jié)果圖長什么樣子：

主要是要看這兩個數(shù)據(jù)集的，這兩個數(shù)據(jù)集長這樣子：

ks值每個模型的ks值，p值是ks的p值，model對應(yīng)的是哪個模型，GLIM是哪個廣義線性模型，1_WD1代表的是隱藏層為1，權(quán)衰減為0.1對應(yīng)的模型，在1_WD1，代表的隱藏層為1的時候?qū)?yīng)的早停止發(fā)的模型，在2_WD1，代表的隱藏層為2的時候?qū)?yīng)的早停止發(fā)的模型，找出你喜歡模型之后，去路徑下面找規(guī)則代碼就可以了。如果實在是這個代碼格式跟你的sas不符的，可以在后臺跟我要下txt的格式的代碼。