Databases Reference
In-Depth Information
relationship, no modeling at all is recommended. The dataset viewers of
VisMiner are a good tool to begin this initial exploration and analysis.
Polynomial Regression
To convert a linear regression modeler into a non-linear modeler, a common
method is to add additional columns that are non-linear transformations of the
original input columns. In a second-order polynomial regression, the inputs are
squared. For example, the simple linear regression model becomes:
2
Y
i
¼
b
0
þ
b
1
X
i
þ
b
2
X
i
þ
e
i
A second-order polynomial regression using two or more numeric input
columns would add squared terms for each.
A third-order polynomial regression becomes:
2
3
Y
i
¼
b
0
þ
b
1
X
i
þ
b
2
X
i
þ
b
3
X
i
þ
e
i
Again, when multiple numeric input columns are included in the model, each
input would be expanded to include both the squared and cubed terms. A similar
expansion is possible for polynomial regressions of even higher order.
Artificial Neural Networks for Regression Analysis
To model more complex relationships beyond that provided by polynomial
regression, artificial neural networks (ANN) are a good choice. Using ANN
processors, models can be constructed to fit almost any relationship. They are
especially good at modeling relationships where input attributes interact. Input
interaction occurs when the levels of one input attribute change the relationship
of a second input attribute with respect to the output attribute. For example, in a
particular setting, it may be that the effect of advertising on sales is different for
men and women. In this case, the input attribute “Gender” changes the pattern
of the contribution of input attribute “Advertising” on output attribute “Sales”.
Dataset Preparation
The dataset used for regression modeling in VisMiner must be in a tabular
format. The input attributes may be of any data type - numeric, ordinal, or
nominal. The output attribute must be numeric.
It is important to remember that when using VisMiner to build a regression
model, the dataset should contain only the attributes (input and output) to be
