The HPNEURAL Procedure

TARGET Statement

TARGET variables </ <LEVEL=INT | LEVEL= NOM >
<ACT=activation-function> <ERROR=error-function>>;

The TARGET statement identifies the variables in the input data set that the network is to be trained to predict.

You can specify the following options:

LEVEL=INT | NOM

specifies the variables type. You can specify the following values:

INT: specifies that the variables are interval variables, which must be numeric.
NOM: specifies that the variables are nominal variables, also known as classification variables, which can be numeric or character.

By default, LEVEL=INT.

ACT=COS | EXP | IDENTITY | TANH

specifies the activation function for interval target variables only. You cannot specify the activation function for nominal variables. It is always the softmax function.

You can specify the following values:

COS: specifies the cosine function.
EXP: specifies the exponential function.
IDENTITY: specifies the identity function.
TANH: specifies the hyperbolic tangent function.

By default, ACT=IDENTITY.

ERROR=GAMMA | NORMAL | POISSON

specifies the error function for interval target variables only. The optimizer uses this function to characterize the difference between the network output and the target value. You can specify the following values:

GAMMA: specifies the gamma function. The gamma error function is usually used when you want to predict the time between events.
NORMAL: specifies the normal function, which is the sum of the squared differences between the network output and the target value.
POISSON: specifies the Poisson function. The Poisson error function is usually used when you want to predict the number of events per unit time.

By default, ERROR=NORMAL.

You can specify the ACT= and ERROR= options only for interval variables. You can specify ACT=EXP if and only if you also specify ERROR=GAMMA or ERROR=POISSON.

Nominal variables have one target neuron per class level, except for nominal variables that have only two levels (binary variables), which have a single neuron. Each of these neurons uses the softmax activation function to ensure that the sum of the outputs for all neurons is 1.0. The output of each neuron can then be interpreted as the probability that the variable is the corresponding class level. The error function for nominal variables is always the cross entropy function. Neither the ACT= option nor the ERROR= option is allowed for nominal variables.

When training, you must include one or more TARGET statements. You need more than one TARGET statement when you have both interval and nominal target variables. The TARGET statement is not allowed when you do stand-alone scoring.

When you are training, any observation that has missing values for any variable is not used.

You cannot specify the same variable in both an INPUT statement and a TARGET statement.

Last updated: January 04, 2019