32
§
PROC DMNEURL: Approximation to PROC NEURAL
WEIGHT or WEIGHTS Statement
WEIGHT onevar ;
WEIGHTS onevar ;
One numeric (interval scaled) variable may be specified as a WEIGHT variable. It is
recommended to specify the WEIGHT variable already in the PROC DMDB invoca-
tion. Then the information is saved in the catalog and that variable is used automati-
cally as a FREQ variable in PROC DMNEURL.
Scoring the Model Using the OUTEST= Data set
The score value
Ý
is computed for each observation
½
Æ
Ó
×
with nonmissing
value of the target (response) variable
Ý
of the input data set. All information needed
for scoring an observation of the DMDB data set is contained in the output of the
OUTEST= data set. First an observation from the input data set is mapped into a
vector
Ú
of
Ò
new values in which
1. CLASS predictor variables with
Ã
categories are replaced by
Ã
·
½
or
Ã
dummy (binary) variables, depending on the fact whether the variable has miss-
ing values or not.
2. Missing values in interval predictor variables are replaced by the mean value of
this variable in the DMDB data set. This mean value is taken from the catalog
of the DMDB data set.
3. The values of a WEIGHT or FREQ variable are multiplied into the observation.
4. For an interval target variable
Ý
its value is transformed into the interval [0,1]
by the relationship
Ý
Ò
Û
Ý
Ý
Ñ
Ò
Ý
Ñ
Ü
Ý
Ñ
Ò
5. All predictor variables are transformed into values with zero mean and unit
standard deviation by
Ü
Ò
Û
Ü
Å
Ò´Ü
µ
Ë
Ø
Ú
´Ü
µ
The values for
Å
Ò´Ü
µ
and
Ë
Ø
Ú
´Ü
µ
are listed in the OUTEST= data set.
This means, that in the presence of CLASS variables the n-vector
Ú
has more entries
than the observation in the data set.
The scoring is additive across the stages. The following information is available for
scoring each stage
¯
components (eigenvectors)
Þ
Ð
each of dimension
Ò
¯
the best activation function
and a specified link function
¯
the
Ô
¾
·
½
optimal parameter estimates
Purpose of PROC DMNEURL
§
33
For each component
Þ
Ð
we compute the component score
Ù
Ð
,
Ù
Ð
Ò
½
Þ
Ð
Ú
similar to principal component analysis. With those values
Ù
Ð
the model can be ex-
pressed as
Ý
Ò×Ø
×Ø
½
´
´Ù
µµ
where
is the best activation function and
is the specified link function.
In other words, this means, that given the
Ù
Ð
the value
Û
is computed from
Û
¼
·
Ð
´Ù
Ð
Ð
Ð
µ
where
Ð
and
Ð
are two of the
Ô
¾
£
·
½
optimal parameters
and
is defined as
SQUARE
Û
´
·
£
Ùµ
£
Ù
TANH
Û
£
Ø
Ò
´
£
Ùµ
ARCTAN
Û
£
Ø
Ò´
£
Ùµ
LOGIST
Û
ÜÔ´
£
Ùµ
´½
·
ÜÔ´
£
Ùµµ
GAUSS
Û
£
ÜÔ´ ´
£
Ùµ
¾
µ
SIN
Û
£
×
Ò´
£
Ùµ
COS
Û
£
Ó×´
£
Ùµ
EXP
Û
£
ÜÔ´
£
Ùµ
For the first component
½
½
and
½
¾
, for the second component
¾
¿
and
¾
, and for the last component
Ô ½
and
Ô
are used.
The link function
is applied on
Û
and yields to
IDENT
Û
LOGIST
ÜÔ´Û
µ
´½
·
ÜÔ´Û
µ
RECIPR
½
Û
Across all stages the values of
are added to the predicted value (posterior)
Ý
.
The DMREG Procedure
The DMREG Procedure
Overview
Procedure Syntax
PROC DMREG Statement
CLASS Statement
CODE Statement
DECISION Statement
FREQ Statement
MODEL Statement
NLOPTIONS Statement
REMOTE Statement
SCORE Statement
Details
Examples
Example 1: Linear and Quadratic Logistic Regression with an Ordinal Target (Rings Data)
Example 2: Performing a Stepwise OLS Regression (DMREG Baseball Data)
Example 3: Comparison of the DMREG and LOGISTIC Procedures when Using a Categorical Input
Variable
References
Copyright 2000 by SAS Institute Inc., Cary, NC, USA. All rights reserved.