The transformation expressions in a transformation record are expression_rec objects.

TYPE expression_rec IS RECORD (
     lstmt       DBMS_SQL.VARCHAR2A,
     lb          BINARY_INTEGER DEFAULT 1,
     ub          BINARY_INTEGER DEFAULT 0);

TYPE varchar2a IS TABLE OF VARCHAR2(32767)
INDEX BY BINARY_INTEGER;

The lstmt field stores a VARCHAR2A, which allows transformation expressions to be very long, as they can be broken up across multiple rows of VARCHAR2. Use the DBMS_DATA_MINING_TRANSFORM.SET_EXPRESSION procedure to create an expression_rec.

The attribute specification in a transformation record defines characteristics that are specific to this attribute. If not null, the attribute specification can include values FORCE_IN, NOPREP, or TEXT, as described in Table 4-2.

The SET_TRANSFORM procedure adds a single transformation record to a transformation list.

DBMS_DATA_MINING_TRANSFORM.SET_TRANSFORM (
          xform_list               IN OUT NOCOPY TRANSFORM_LIST,
          attribute_name           VARCHAR2,
          attribute_subname        VARCHAR2,
          expression               VARCHAR2,
          reverse_expression       VARCHAR2,
          attribute_spec           VARCHAR2 DEFAULT NULL);

SQL expressions that you specify with SET_TRANSFORM must fit within a VARCHAR2. To specify a longer expression, you can use the SET_EXPRESSION procedure, which builds an expression by appending rows to a VARCHAR2 array.

These two functions can be used to create a new transformation list from the transformations embedded in an existing model.

The GET_MODEL_TRANSFORMATIONS function returns a list of embedded transformations.

DBMS_DATA_MINING.GET_MODEL_TRANSFORMATIONS (
      model_name     IN VARCHAR2)
RETURN DM_TRANSFORMS PIPELINED;

GET_MODEL_TRANSFORMATIONS returns a table of dm_transform objects. Each dm_transform has these fields

attribute_name       VARCHAR2(4000)
attribute_subname    VARCHAR2(4000)
expression           CLOB
reverse_expression   CLOB

The components of a transformation list are transform_rec, not dm_transform. The fields of a transform_rec are described in Table 4-2. You can call GET_MODEL_TRANSFORMATIONS to convert a list of dm_transform objects to transform_rec objects and append each transform_rec to a transformation list.

DBMS_DATA_MINING.GET_TRANSFORM_LIST (
      xform_list           OUT NOCOPY TRANSFORM_LIST,
      model_xforms         IN  DM_TRANSFORMS);

A number of factors go into deciding a binning strategy. Having fewer values typically leads to a more compact model and one that builds faster, but it can also lead to some loss in accuracy.

Model quality can improve significantly with well-chosen bin boundaries. For example, an appropriate way to bin ages is to separate them into groups of interest, such as children 0-13, teenagers 13-19, youth 19-24, working adults 24-35, and so on.

The following table lists the binning techniques provided by Oracle Data Mining:

Most normalization methods map the range of a single attribute to another range, typically 0 to 1 or -1 to +1.

Normalization is very sensitive to outliers. Without outlier treatment, most values are mapped to a tiny range, resulting in a significant loss of information.

Outliers are extreme values, typically several standard deviations from the mean. To minimize the effect of outliers, you can Winsorize or trim the data.

Winsorizing involves setting the tail values of an attribute to some specified value. For example, for a 90% Winsorization, the bottom 5% of values are set equal to the minimum value in the 5th percentile, while the upper 5% of values are set equal to the maximum value in the 95th percentile.

Trimming sets the tail values to NULL. The algorithm treats them as missing values.

Outliers affect the different algorithms in different ways. In general, outliers cause distortion with equi-width binning and min-max normalization.