You can use the neural net algorithm for both prediction and classification.
This algorithm is much more powerful and flexible than linear regression. For example, you can specify multiple targets as well as multiple predictors.
On the other hand, the model generated by the neural net algorithm is not as easy to interpret as the model generated from linear regression.
The following table shows the parameters for the neural net algorithm.
Parameter Name | Description |
|---|---|
Activation specifies the activation function | The activation function for the algorithm. Choose one from the function list. Default: Logistic Sigmoid |
HiddenLayerSize specifies the number of nodes in the hidden layer. | This parameter determines the speed and accuracy of the algorithm. In general, set this parameter to the square root of the product of the number of predictor accessors and target accessors. Values that are too high consume training time and produce unpredictable results. Values that are too low produce inaccurate results. Type: integer Example: If you have four predictors and two targets, set this parameter to 3 (square root of (4 * 2) = 2.8, rounded up to 3). Default: 2 |
PredictorRange specifies the half range of the predictor data. | This parameter specifies the expected half range of the predictor accessor data. Type: double Example: A setting of 1000 means that the data ranges from -1,000 to 1,000. Default: 1,000,000 |
TargetRange specifies the half range of the target data. | This parameter specifies the expected half range of the target accessor data. Type: double Example: A setting of 1000 means that the data ranges from -1,000 to 1,000. Default: 1,000,000 |
You can enter multiple predictor and target attributes, all of which are numeric. The number of predictor attributes must exceed the number of target attributes.
You must normalize predictor and target values to values between 0 and 1. The use of denormalized values can affect the convergence of the net or cause other numerical problems.
One use of neural nets is binary classification. A series of inputs (predictors) produces a set of results (targets) normalized to values between 0 and 1. For example, a set of behaviors results in values between 0 and 1, with 1 being risky and 0 being risk free. Values in between require interpretation; for example, 0.15 is the high end of safe and 0.85 is the low end of risky.
The following tables show the required and optional predictors and targets for the neural net algorithm.
Table 10. Neural Net Algorithm Predictors
Member Expression | Sample |
|---|---|
Predictor.Predictor specifies member or member set to use for the predictor domain. | {[Cream_Soda].Children} |
Predictor.Sequence defines the sequence to be traversed for the predictor, generally a time dimension range | {[Jan 1].Level.Members} |
Predictor.External (optional) defines the scope of the predictor | (none provided) |
Predictor.Anchor (optional) specifies additional restrictions from other dimensions. | {([2001], [Actual], [Sales], [Houston])} |
Table 11. Neural Net Algorithm Targets
Target Expression | Sample |
|---|---|
Target.Target specifies member set to use for the target domain. | {[Cola].Children} |
Target.Sequence defines the sequence to be traversed for the target, generally a time dimension range. | {[Jan 1].Level.Members} |
Target.External (optional) defines the scope of the target | (none provided) |
Target.Anchor (optional) specifies additional restrictions from other dimensions. | {([2001], [Actual], [Sales], [Houston])} |
This information will be made available at a later release.
This information will be made available at a later release.