42 DBMS_DATA_MINING

Using DBMS_DATA_MINING

This section contains topics that relate to using the DBMS_DATA_MINING package.

• Overview

• Mining Model Objects

• Deprecated Subprograms

• Mining Functions

• Model Settings

• Data Types

Overview

Oracle Data Mining supports both supervised and unsupervised data mining. Supervised data mining predicts a target value based on historical data. Unsupervised data mining discovers natural groupings and does not use a target.

A data mining function refers to the methods for solving a given class of data mining problems. The mining function must be specified when a model is created. See "Mining Functions".

Supervised data mining functions include:

• Classification

• Regression

• Attribute Importance

Unsupervised data mining functions include:

• Clustering

• Association

• Feature Extraction

• Anomaly Detection (one-class classification)

The steps you use to build and apply a mining model depend on the data mining function and the algorithm being used. The algorithms supported by Oracle Data Mining are listed in Table 42-1.

Mining Model Objects

Mining models are Oracle Database schema objects. They support the standard security features of Oracle Database. Mining models are also supported by SQL COMMENT and SQL AUDIT.

ALL_MINING_MODELS

You can query the data dictionary view ALL_MINING_MODELS to obtain a list of accessible mining models.

SQL> describe all_mining_models
 Name                                      Null?    Type
 ----------------------------------------- -------- ----------------------------
 OWNER                                     NOT NULL VARCHAR2(30)
 MODEL_NAME                                NOT NULL VARCHAR2(30)
 MINING_FUNCTION                                    VARCHAR2(30)
 ALGORITHM                                          VARCHAR2(30)
 CREATION_DATE                             NOT NULL DATE
 BUILD_DURATION                                     NUMBER
 MODEL_SIZE                                         NUMBER
 COMMENTS                                           VARCHAR2(4000)

Mining Model Naming Restrictions

The naming rules for models are more restrictive than the naming rules for most database schema objects. A model name must satisfy the following additional requirements:

• It must be 25 or fewer characters long.

• It must be a nonquoted identifier. Oracle requires that nonquoted identifiers contain only alphanumeric characters, the underscore (_), dollar sign ($), and pound sign (#); the initial character must be alphabetic. Oracle strongly discourages the use of the dollar sign and pound sign in nonquoted literals.

Naming requirements for schema objects are fully documented in Oracle Database SQL Language Reference.

ALL_MINING_MODEL_ATTRIBUTES

You can query the data dictionary view ALL_MINING_MODEL_ATTRIBUTES to obtain a list of the data attributes for each accessible mining model. Data attributes are the columns of data used by an algorithm to build a model. Some or all of these columns must be present in the data to which the model is applied.

Data attributes are referred to as the model signature. The ALL_MINING_MODEL_ATTRIBUTES view lists the data attributes in the model signature, including the target if the model is supervised.

An algorithm builds an internal representation of the data attributes and uses them as either categoricals (data that classifies or categorizes) or as numericals (continuous data). These internal model attributes can be viewed using the GET_MODEL_DETAILS functions.

SQL> describe all_mining_model_attributes
 Name                                      Null?    Type
 ----------------------------------------- -------- ----------------------------
 OWNER                                     NOT NULL VARCHAR2(30)
 MODEL_NAME                                NOT NULL VARCHAR2(30)
 ATTRIBUTE_NAME                            NOT NULL VARCHAR2(30)
 ATTRIBUTE_TYPE                                     VARCHAR2(11)
 DATA_TYPE                                          VARCHAR2(12)
 DATA_LENGTH                                        NUMBER
 DATA_PRECISION                                     NUMBER
 DATA_SCALE                                         NUMBER
 USAGE_TYPE                                         VARCHAR2(8)
 TARGET                                             VARCHAR2(3)

ALL_MINING_MODEL_SETTINGS

The view ALL_MINING_MODEL_SETTINGS returns the settings for each accessible mining model. Settings control various characteristics of mining models.

All settings have default values. The values of some settings are generated by the algorithm by default. You can override the default value of a setting by specifying its value in a settings table for the model. All settings, both default and user-specified, are listed in ALL_MINING_MODEL_SETTINGS.

SQL> describe all_mining_model_settings
 Name                                      Null?    Type
 ----------------------------------------- -------- ----------------------------
 OWNER                                     NOT NULL VARCHAR2(30)
 MODEL_NAME                                NOT NULL VARCHAR2(30)
 SETTING_NAME                              NOT NULL VARCHAR2(30)
 SETTING_VALUE                                      VARCHAR2(4000)
 SETTING_TYPE                                       VARCHAR2(7)

Deprecated Subprograms

The following subprograms are deprecated in 11g Release 2 (11.2).

• GET_DEFAULT_SETTINGS

  Replaced with data dictionary views: USER/ALL/DBA_MINING_MODEL_SETTINGS

• GET_MODEL_SETTINGS

  Replaced with data dictionary views: USER/ALL/DBA_MINING_MODEL_SETTINGS

• GET_MODEL_SIGNATURE

  Replaced with data dictionary views: USER/ALL/DBA_MINING_MODEL_ATTRIBUTES

The following view is deprecated in 11g Release 2 (11.2).

• DM_USER_MODELS

  Replaced with data dictionary views: USER/ALL/DBA_MINING_MODELS

The Adaptive Bayes Network algorithm is deprecated in 11g Release 2 (11.2).

The DMSYS schema no longer exists in 11g Release 2 (11.2). Oracle Data Mining metadata is now in SYS.

Mining Functions

The constants that specify the mining function of a model are listed in Table 42-2. The concept of a "mining function" is introduced in "Overview".

All models are created with a mining function. The mining function is a required argument to the CREATE_MODEL Procedure.

Model Settings

Oracle Data Mining uses settings to specify the algorithm and other characteristics of a model. Some settings are general, some are specific to a mining function, and some are specific to an algorithm.

All settings have default values. If you want to override one or more of the settings for a model, you must create a settings table. The settings table must have the column names and data types shown in Table 42-3.

The information you provide in the settings table is used by the model at build time. The name of the settings table is an optional argument to the CREATE_MODEL Procedure.

You can find the settings used by a model by querying the data dictionary view ALL_MINING_MODEL_SETTINGS. This view lists the model settings used by the mining models to which you have access. All the setting values are included in the view, whether default or user-specified. See "ALL_MINING_MODEL_SETTINGS".

Algorithm Names

The ALGO_NAME setting specifies the model algorithm. The values for the ALGO_NAME setting are listed in Table 42-4.

Oracle Data Mining supports more than one algorithm for the classification, regression, and clustering mining functions. Each of these mining functions has a default algorithm, as shown in Table 42-5.

Automatic Data Preparation

The PREP_AUTO setting indicates whether or not the model will use Automatic Data Preparation (ADP). By default ADP is disabled.

When you enable ADP, the model uses heuristics to transform the build data according to the requirements of the algorithm. The transformation instructions are stored with the model and reused whenever the model is applied. You can view the transformation instructions in the model details.

You can choose to supplement automatic data preparations by specifying additional transformations in the xform_list parameter when you build the model. (See "CREATE_MODEL Procedure".)

If you do not use ADP (default) and do not specify transformations in the xform_list parameter to CREATE_MODEL (also the default), you will continue to operate in 10.2 mode. This means that you must implement your own transformations separately in the build, test, and scoring data; you must take special care to implement the exact same transformations in each data set.

If you do not use ADP, but you do specify transformations in the xform_list parameter to CREATE_MODEL, Oracle Data Mining embeds the transformation definitions in the model and prepares the test and scoring data to match the build data. Because of automatic and embedded data preparation, mining models are known as supermodels.

The values for the PREP_AUTO setting are described in Table 42-6.

Mining Function Settings

The settings described in Table 42-7 apply to a mining function.

Global Settings

The settings in Table 42-8 are applicable to any type of model, but are currently only implemented for specific algorithms.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_MINING_MODEL_SETTINGS.

Algorithm Settings: Adaptive Bayes Network (deprecated)

These settings affect the behavior of the Adaptive Bayes Network algorithm.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: Decision Tree

These settings affect the behavior of the Decision Tree algorithm.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: Generalized Linear Models

These settings affect the behavior of GLM models. GLM can be used for classification (logistic regression) or regression (linear regression).

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: k-Means

These settings affect the behavior of the k-Means algorithm.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: Naive Bayes

These settings affect the behavior of the Naive Bayes Algorithm.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: Non-Negative Matrix Factorization

These settings affect the behavior of the Non-Negative Matrix Factorization algorithm.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: O-Cluster

These settings affect the behavior of the O-Cluster algorithm.

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Algorithm Settings: Support Vector Machine

These settings affect the behavior of the Support Vector Machine algorithm. SVM can be used for classification or regression, or for anomaly detection (classification with a null target).

You can query the data dictionary view *_MINING_MODEL_SETTINGS (using the ALL, USER, or DBA prefix) to find the setting values for a model. See Oracle Data Mining Application Developer's Guide for information about *_ALL_MINING_MODEL_SETTINGS.

Data Types

The DBMS_DATA_MINING package uses object data types to store information about model attributes. Most of these types are returned by the table functions GET_n, where n identifies the type of information to return. These functions take a model name as input and return the requested information as a collection of rows.

For a list of the GET functions, see "Summary of DBMS_DATA_MINING Subprograms".

Oracle Data Mining also uses object data types for handling transactional data. These types, DM_NESTED_NUMERICALS and DM_NESTED_CATEGORICALS specify nested tables that can be used for storing a set of mining attributes in a single column. For more information on nested tables, see the Oracle Data Mining Application Developer's Guide.

All the table functions use pipelining, which causes each row of output to be materialized as it is read from model storage, without waiting for the generation of the complete table object. For more information on pipelined, parallel table functions, consult the Oracle Database PL/SQL Language Reference.

The Data Mining object data types are described in Table 42-17.

Summary of DBMS_DATA_MINING Subprograms

Table 42-18 summarizes the subprograms included in the DBMS_DATA_MINING package.

ADD_COST_MATRIX Procedure

This procedure associates a cost matrix table with a classification model. The cost matrix biases the model by assigning costs or benefits to specific model outcomes.

The cost matrix is stored with the model and taken into account when the model is scored. The stored cost matrix is the default scoring matrix for the model.

You can also specify a cost matrix inline when you invoke a Data Mining SQL function for scoring. When an inline cost matrix is specified, it is used instead of the default, stored cost matrix (if one exists).

To obtain the default scoring matrix for a model, use the GET_MODEL_COST_MATRIX function. To remove the default scoring matrix from a model, use the REMOVE_COST_MATRIX procedure. See "GET_MODEL_COST_MATRIX Function" and "REMOVE_COST_MATRIX Procedure".

Syntax

DBMS_DATA_MINING.ADD_COST_MATRIX (
       model_name                IN VARCHAR2,
       cost_matrix_table_name    IN VARCHAR2,
       cost_matrix_schema_name   IN VARCHAR2 DEFAULT NULL);

Parameters

Usage Notes

1. If the model is not in your schema, then ADD_COST_MATRIX requires the ALTER ANY MINING MODEL system privilege or the ALTER object privilege for the mining model.

2. The cost matrix table must have the columns shown in Table 42-20. Note that the actual and predicted target values must have the same data type.

   Table 42-20 Required Columns in a Cost Matrix Table

   Column Name	Data Type
   ACTUAL_TARGET_VALUE	VARCHAR2(4000) for categorical targets NUMBER for numeric targets
   PREDICTED_TARGET_VALUE	VARCHAR2(4000)for categorical targets NUMBER for numeric targets
   COST	NUMBER

   
   

3. Since a benefit can be viewed as a negative cost, you can specify a benefit for a given outcome by providing a negative number in the costs column of the cost matrix table.

4. All classification algorithms can use a cost matrix for scoring. The Decision Tree algorithm can also use a cost matrix at build time.If you want to build a Decision Tree model with a cost matrix, specify the cost matrix table name in the CLAS_COST_TABLE_NAME setting in the settings table for the model. See Table 42-7, "Mining Function Settings".

   The cost matrix used to create a Decision Tree model becomes the default scoring matrix for the model. If you want to specify different costs for scoring, use the REMOVE_COST_MATRIX procedure to remove the cost matrix and the ADD_COST_MATRIX procedure to add a new one.

Example

This example creates a cost matrix table called COSTS_NB and adds it to a Naive Bayes model called NB_SH_CLAS_SAMPLE. The model has a binary target: 1 means that the customer responds to a promotion; 0 means that the customer does not respond. The cost matrix assigns a cost of .25 to misclassifications of customers who do not respond and a cost of .75 to misclassifications of customers who do respond. This means that it is three times more costly to misclassify responders than it is to misclassify non-responders.

CREATE TABLE costs_nb (
  actual_target_value           NUMBER,
  predicted_target_value        NUMBER,
  cost                          NUMBER);
INSERT INTO costs_nb values (0, 0, 0);
INSERT INTO costs_nb values (0, 1, .25);
INSERT INTO costs_nb values (1, 0, .75);
INSERT INTO costs_nb values (1, 1, 0);
COMMIT;
 
EXEC dbms_data_mining.add_cost_matrix('nb_sh_clas_sample', 'costs_nb');
 
SELECT cust_gender, COUNT(*) AS cnt, ROUND(AVG(age)) AS avg_age
   FROM mining_data_apply_v
   WHERE PREDICTION(nb_sh_clas_sample COST MODEL
      USING cust_marital_status, education, household_size) = 1
   GROUP BY cust_gender
   ORDER BY cust_gender;
   
C        CNT    AVG_AGE
- ---------- ----------
F         72         39
M        555         44

ALTER_REVERSE_EXPRESSION Procedure

This procedure changes a reverse transformation expression to an expression that you specify.

A transformation expression specifies how an attribute will be changed before it is used to create a model. The reverse transformation expression specifies how the attribute will be returned to its original form (or some approximation of its original form) for purposes of model transparency.

The GET_MODEL_DETAILS functions apply reverse transformations before returning information about a model's attributes. When the target is transformed, its reverse transformation is applied to the results of scoring. See the Usage Notes.

You can use ALTER_REVERSE_EXPRESSION to associate a name/description with a cluster or feature; the descriptor that you specify will be displayed when the model is scored.You can use the CLUSTER_ID, CLUSTER_PROBABILITY, and CLUSTER_SET functions to score clustering models. You can use the FEATURE_ID, FEATURE_SET, or FEATURE_VALUE functions to score feature extraction models. See the example and Oracle Database SQL Language Reference.

You can use ALTER_REVERSE_EXPRESSION to associate a name/description with the results of an anomaly detection model. These models use One-Class SVM to predict either 1 or 0 for each record, indicating whether or not the record is anomalous. With ALTER_REVERSE_EXPRESSION, you can replace the zeros and ones with labels that are more meaningful.

Syntax

DBMS_DATA_MINING. ALTER_REVERSE_EXPRESSION (
         model_name             VARCHAR2,
         expression             CLOB,
         attribute_name         VARCHAR2 DEFAULT NULL,
         attribute_subname      VARCHAR2 DEFAULT NULL);

Parameters

Usage Notes

If you are using automatic data preparation, the transformations and reverse transformations are performed automatically by the system. When you create the model, you can specify transformations and reverse transformations to be applied in addition to the system-generated expressions. If you are not using automatic data preparation, the transformations and reverse transformations that you specify are the only ones that are applied.

Example

This example specifies labels for the clusters that result from the sh_clus model. The labels consist of the word "Cluster" and the internal numeric identifier for the cluster.

SQL> begin
  2    dbms_data_mining.alter_reverse_expression(
  3      'sh_clus', '''Cluster ''||value');
  4  end;
  5  /
 
SQL> select cust_id, cluster_id(sh_clus using *) cluster_id
  2    from sh_aprep_num
  3   where cust_id < 100011
  4   order by cust_id;
 
CUST_ID CLUSTER_ID
------- ------------------------------------------------
 100001 Cluster 18
 100002 Cluster 14
 100003 Cluster 14
 100004 Cluster 18
 100005 Cluster 19
 100006 Cluster 7
 100007 Cluster 18
 100008 Cluster 14
 100009 Cluster 8
 100010 Cluster 8

APPLY Procedure

This procedure applies a mining model to the data of interest, and generates the results in a table. The apply process is also referred to as scoring.

For predictive mining functions, the apply process generates predictions in a target column. For descriptive mining functions such as clustering, the apply process assigns each case to a cluster with a probability.

In Oracle Data Mining, the apply operation is not applicable to association models and attribute importance models.

Syntax

DBMS_DATA_MINING.APPLY (
      model_name           IN VARCHAR2,
      data_table_name      IN VARCHAR2,
      case_id_column_name  IN VARCHAR2,
      result_table_name    IN VARCHAR2,
      data_schema_name     IN VARCHAR2 DEFAULT NULL);

Parameters

Usage Notes

1. The data provided for APPLY must undergo the same preprocessing as the data used to create and test the model. When you use Automatic Data Preparation, the preprocessing required by the algorithm is handled for you by the model — both at build time and apply time. (See "Automatic Data Preparation".)

2. APPLY creates a table in the user's schema to hold the results. The columns are algorithm-specific.

   The columns in the results table are listed in Table 42-23 through Table 42-27. The case ID column name in the results table will match the case ID column name provided by you. The type of the incoming case ID column is also preserved in APPLY output.

   Note:

   Make sure that the case ID column does not have the same name as one of the columns that will be created by APPLY. For example, when applying a classification model, the case ID in the scoring data must not be 'PREDICTION' or 'PROBABILITY' (See Table 42-23).

3. The data type for the 'PREDICTION', 'CLUSTER_ID', and 'FEATURE_ID' output columns is influenced by any reverse expression that is embedded in the model by the user. If the user does not provide a reverse expression that alters the scored value type, then the types will conform to the descriptions in the following tables. See "ALTER_REVERSE_EXPRESSION Procedure".

Classification

The results table for classification has the columns described in Table 42-23. If the target of the model is categorical, the PREDICTION column will have a VARCHAR2 data type. If the target is numerical, the PREDICTION column will have a NUMBER data type.

One-Class SVM (Anomaly Detection)

The results table for anomaly detection has the columns described in Table 42-24.

Values in the PREDICTION column can be either 0 or 1. When the prediction is 1, the case is a typical example. When the prediction is 0, the case is an outlier.

Regression using SVM or GLM

The results table for regression has the columns described in Table 42-25.

Clustering using k-Means or O-Cluster

Clustering is an unsupervised mining function, and hence there are no targets. The results of an APPLY operation will contain simply the cluster identifier corresponding to a case, and the associated probability. The results table has the columns described in Table 42-26.

Feature Extraction using NMF

Feature extraction is also an unsupervised mining function, and hence there are no targets. The results of an APPLY operation will contain simply the feature identifier corresponding to a case, and the associated match quality. The results table has the columns described in Table 42-27.

Examples

This example applies the GLM regression model GLMR_SH_REGR_SAMPLE to the data in the MINING_DATA_APPLY_V view. The apply results are output to the table REGRESSION_APPLY_RESULT.

SQL> BEGIN
       DBMS_DATA_MINING.APPLY (
       model_name     => 'glmr_sh_regr_sample',
       data_table_name     => 'mining_data_apply_v',
       case_id_column_name => 'cust_id',
       result_table_name   => 'regression_apply_result');
    END;
    /
 
SQL> SELECT * FROM regression_apply_result WHERE cust_id >  101485;
 
   CUST_ID PREDICTION
---------- ----------
    101486 22.8048824
    101487 25.0261101
    101488 48.6146619
    101489   51.82595
    101490 22.6220714
    101491 61.3856816
    101492 24.1400748
    101493  58.034631
    101494 45.7253149
    101495 26.9763318
    101496 48.1433425
    101497 32.0573434
    101498 49.8965531
    101499  56.270656
    101500 21.1153047

COMPUTE_CONFUSION_MATRIX Procedure

This procedure computes a confusion matrix, stores it in a table in the user's schema, and returns the model accuracy.

A confusion matrix is a test metric for classification models. It compares the predictions generated by the model with the actual target values in a set of test data. The matrix is n-by-n, where n is the number of classes. The confusion matrix lists the number of times each class was correctly predicted and the number of times it was predicted to be one of the other classes.

COMPUTE_CONFUSION_MATRIX accepts three input streams:

• The predictions generated on the test data. The information is passed in three columns:

  • Case ID column

  • Prediction column

  • Scoring criterion column containing either probabilities or costs

• The known target values in the test data. The information is passed in two columns:

  • Case ID column

  • Target column containing the known target values

• (Optional) A cost matrix table with predefined columns. See the Usage Notes for the column requirements.

Syntax

DBMS_DATA_MINING.COMPUTE_CONFUSION_MATRIX (
      accuracy                     OUT NUMBER,
      apply_result_table_name      IN  VARCHAR2,
      target_table_name            IN  VARCHAR2,
      case_id_column_name          IN  VARCHAR2,
      target_column_name           IN  VARCHAR2,
      confusion_matrix_table_name  IN  VARCHAR2,
      score_column_name            IN  VARCHAR2 DEFAULT 'PREDICTION',
      score_criterion_column_name  IN  VARCHAR2 DEFAULT 'PROBABILITY',
      cost_matrix_table_name       IN  VARCHAR2 DEFAULT NULL,
      apply_result_schema_name     IN  VARCHAR2 DEFAULT NULL,
      target_schema_name           IN  VARCHAR2 DEFAULT NULL,
      cost_matrix_schema_name      IN  VARCHAR2 DEFAULT NULL,
      score_criterion_type         IN  VARCHAR2 DEFAULT 'PROBABILITY');

Parameters

Usage Notes

• The predictive information you pass to COMPUTE_CONFUSION_MATRIX may be generated using SQL PREDICTION functions, the DBMS_DATA_MINING.APPLY procedure, or some other mechanism. As long as you pass the appropriate data, the procedure can compute the confusion matrix.

• Instead of passing a cost matrix to COMPUTE_CONFUSION_MATRIX, you can use a scoring cost matrix associated with the model. A scoring cost matrix can be embedded in the model or it can be defined dynamically when the model is applied. To use a scoring cost matrix, invoke the SQL PREDICTION_COST function to populate the score criterion column.

• The predictions that you pass to COMPUTE_CONFUSION_MATRIX are in a table or view specified in apply_result_table_name.

  CREATE TABLE apply_result_table_name AS (
              case_id_column_name            VARCHAR2, 
              score_column_name              VARCHAR2,
              score_criterion_column_name    VARCHAR2);
  

• A cost matrix must have the columns described in Table 42-29.

  Table 42-29 Columns in a Cost Matrix

  Column Name	Data Type
  actual_target_value	NUMBER or VARCHAR2
  predicted_target_value	NUMBER or VARCHAR2
  cost	NUMBER

  
  

  See Also:

  Oracle Data Mining Concepts for more information about cost matrixes

• The confusion matrix created by COMPUTE_CONFUSION_MATRIX has the columns described in Table 42-30.

  Table 42-30 Columns in a Confusion Matrix

  Column Name	Data Type
  actual_target_value	NUMBER or VARCHAR2
  predicted_target_value	NUMBER or VARCHAR2
  value	NUMBER

  
  

  See Also:

  Oracle Data Mining Concepts for more information about confusion matrixes

Examples

These examples use the Naive Bayes model nb_sh_clas_sample, which is created by one of the Oracle Data Mining sample programs.

Compute a Confusion Matrix Based on Probabilities

The following statement applies the model to the test data and stores the predictions and probabilities in a table.

CREATE TABLE nb_apply_results AS
       SELECT cust_id,
              PREDICTION(nb_sh_clas_sample USING *) prediction,
              PREDICTION_PROBABILITY(nb_sh_clas_sample USING *) probability
       FROM mining_data_test_v;

Using probabilities as the scoring criterion, you can compute the confusion matrix as follows.

DECLARE
   v_accuracy    NUMBER;
      BEGIN
        DBMS_DATA_MINING.COMPUTE_CONFUSION_MATRIX (
                   accuracy                     => v_accuracy,
                   apply_result_table_name      => 'nb_apply_results',
                   target_table_name            => 'mining_data_test_v',
                   case_id_column_name          => 'cust_id',
                   target_column_name           => 'affinity_card',
                   confusion_matrix_table_name  => 'nb_confusion_matrix',
                   score_column_name            => 'PREDICTION',
                   score_criterion_column_name  => 'PROBABILITY'
                   cost_matrix_table_name       =>  null,                   apply_result_schema_name     =>  null,                   target_schema_name           =>  null,                   cost_matrix_schema_name      =>  null,                   score_criterion_type         => 'PROBABILITY');
        DBMS_OUTPUT.PUT_LINE('**** MODEL ACCURACY ****: ' || ROUND(v_accuracy,4));
      END;
      /

The confusion matrix and model accuracy are shown as follows.

**** MODEL ACCURACY ****: .7847

SQL>SELECT * from nb_confusion_matrix;
ACTUAL_TARGET_VALUE PREDICTED_TARGET_VALUE      VALUE
------------------- ---------------------- ----------
                  1                      0         60
                  0                      0        891
                  1                      1        286
                  0                      1        263

Compute a Confusion Matrix Based on a Cost Matrix Table

The confusion matrix in the previous example shows a high rate of false positives. For 263 cases, the model predicted 1 when the actual value was 0. You could use a cost matrix to minimize this type of error.

The cost matrix table nb_cost_matrix specifies that a false positive is 3 times more costly than a false negative.

SQL> SELECT * from nb_cost_matrix;
ACTUAL_TARGET_VALUE PREDICTED_TARGET_VALUE       COST
------------------- ---------------------- ----------
                  0                      0          0
                  0                      1        .75
                  1                      0        .25
                  1                      1          0

This statement shows how to generate the predictions using APPLY.

BEGIN
    DBMS_DATA_MINING.APPLY(
          model_name          => 'nb_sh_clas_sample',
          data_table_name     => 'mining_data_test_v',
          case_id_column_name => 'cust_id',
          result_table_name   => 'nb_apply_results');
 END;
/

This statement computes the confusion matrix using the cost matrix table. The score criterion column is named 'PROBABILITY', which is the name generated by APPLY.

DECLARE
  v_accuracy    NUMBER;
     BEGIN
       DBMS_DATA_MINING.COMPUTE_CONFUSION_MATRIX (
                accuracy                     => v_accuracy,
                apply_result_table_name      => 'nb_apply_results',
                target_table_name            => 'mining_data_test_v',
                case_id_column_name          => 'cust_id',
                target_column_name           => 'affinity_card',
                confusion_matrix_table_name  => 'nb_confusion_matrix',
                score_column_name            => 'PREDICTION',
                score_criterion_column_name  => 'PROBABILITY',
                cost_matrix_table_name       => 'nb_cost_matrix',
                apply_result_schema_name     => null,
                target_schema_name           => null,
                cost_matrix_schema_name      => null,
                score_criterion_type         => 'COST');
       DBMS_OUTPUT.PUT_LINE('**** MODEL ACCURACY ****: ' || ROUND(v_accuracy,4));
    END;
    /

The resulting confusion matrix shows a decrease in false positives (212 instead of 263).

**** MODEL ACCURACY ****: .798

SQL> SELECT * FROM nb_confusion_matrix;
ACTUAL_TARGET_VALUE PREDICTED_TARGET_VALUE      VALUE
------------------- ---------------------- ----------
                  1                      0         91
                  0                      0        942
                  1                      1        255
                  0                      1        212

Compute a Confusion Matrix Based on Embedded Costs

You can use the ADD_COST_MATRIX procedure to embed a cost matrix in a model. The embedded costs can be used instead of probabilities for scoring. This statement adds the previously-defined cost matrix to the model.

BEGIN    DBMS_DATA_MINING.ADD_COST_MATRIX ('nb_sh_clas_sample', 'nb_cost_matrix');END;/

The following statement applies the model to the test data using the embedded costs and stores the results in a table.

CREATE TABLE nb_apply_results AS
         SELECT cust_id,
              PREDICTION(nb_sh_clas_sample COST MODEL USING *) prediction,
              PREDICTION_COST(nb_sh_clas_sample COST MODEL USING *) cost
          FROM mining_data_test_v;

You can compute the confusion matrix using the embedded costs.

DECLARE
   v_accuracy         NUMBER;
   BEGIN
       DBMS_DATA_MINING.COMPUTE_CONFUSION_MATRIX (
            accuracy                     => v_accuracy,
            apply_result_table_name      => 'nb_apply_results',
            target_table_name            => 'mining_data_test_v',
            case_id_column_name          => 'cust_id',
            target_column_name           => 'affinity_card',
            confusion_matrix_table_name  => 'nb_confusion_matrix',
            score_column_name            => 'PREDICTION',
            score_criterion_column_name  => 'COST',
            cost_matrix_table_name       => null,
            apply_result_schema_name     => null,
            target_schema_name           => null,
            cost_matrix_schema_name      => null,
            score_criterion_type         => 'COST');
   END;
   /

The results are:

**** MODEL ACCURACY ****: .798

SQL> SELECT * FROM nb_confusion_matrix;
ACTUAL_TARGET_VALUE PREDICTED_TARGET_VALUE      VALUE
------------------- ---------------------- ----------
                  1                      0         91
                  0                      0        942
                  1                      1        255
                  0                      1        212

COMPUTE_LIFT Procedure

This procedure computes lift and stores the results in a table in the user's schema.

Lift is a test metric for binary classification models. To compute lift, one of the target values must be designated as the positive class. COMPUTE_LIFT compares the predictions generated by the model with the actual target values in a set of test data. Lift measures the degree to which the model's predictions of the positive class are an improvement over random chance.

Lift is computed on scoring results that have been ranked by probability (or cost) and divided into quantiles. Each quantile includes the scores for the same number of cases.

COMPUTE_LIFT calculates quantile-based and cumulative statistics. The number of quantiles and the positive class are user-specified. Additionally, COMPUTE_LIFT accepts three input streams:

• The predictions generated on the test data. The information is passed in three columns:

  • Case ID column

  • Prediction column

  • Scoring criterion column containing either probabilities or costs associated with the predictions

• The known target values in the test data. The information is passed in two columns:

  • Case ID column

  • Target column containing the known target values

• (Optional) A cost matrix table with predefined columns. See the Usage Notes for the column requirements.

Syntax

DBMS_DATA_MINING.COMPUTE_LIFT (
      apply_result_table_name      IN VARCHAR2,
      target_table_name            IN VARCHAR2,
      case_id_column_name          IN VARCHAR2,
      target_column_name           IN VARCHAR2,
      lift_table_name              IN VARCHAR2,
      positive_target_value        IN VARCHAR2,
      score_column_name            IN VARCHAR2 DEFAULT 'PREDICTION',
      score_criterion_column_name  IN VARCHAR2 DEFAULT 'PROBABILITY',
      num_quantiles                IN NUMBER DEFAULT 10,
      cost_matrix_table_name       IN VARCHAR2 DEFAULT NULL,
      apply_result_schema_name     IN VARCHAR2 DEFAULT NULL,
      target_schema_name           IN VARCHAR2 DEFAULT NULL,
      cost_matrix_schema_name      IN VARCHAR2 DEFAULT NULL
      score_criterion_type         IN VARCHAR2 DEFAULT 'PROBABILITY');

Parameters

Usage Notes

• The predictive information you pass to COMPUTE_LIFT may be generated using SQL PREDICTION functions, the DBMS_DATA_MINING.APPLY procedure, or some other mechanism. As long as you pass the appropriate data, the procedure can compute the lift.

• Instead of passing a cost matrix to COMPUTE_LIFT, you can use a scoring cost matrix associated with the model. A scoring cost matrix can be embedded in the model or it can be defined dynamically when the model is applied. To use a scoring cost matrix, invoke the SQL PREDICTION_COST function to populate the score criterion column.

• The predictions that you pass to COMPUTE_LIFT are in a table or view specified in apply_results_table_name.

  CREATE TABLE apply_result_table_name AS (
              case_id_column_name            VARCHAR2, 
              score_column_name              VARCHAR2,
              score_criterion_column_name    VARCHAR2);
  

• A cost matrix must have the columns described in Table 42-32.

  Table 42-32 Columns in a Cost Matrix

  Column Name	Data Type
  actual_target_value	NUMBER or VARCHAR2
  predicted_target_value	NUMBER or VARCHAR2
  cost	NUMBER

  
  

  See Also:

  Oracle Data Mining Concepts for more information about cost matrixes

• The table created by COMPUTE_LIFT has the columns described in Table 42-33

  Table 42-33 Columns in a Lift Table

  Column Name	Data Type
  quantile_number	NUMBER
  probability_threshold	NUMBER
  gain_cumulative	NUMBER
  quantile_total_count	NUMBER
  quantile_target_count	NUMBER
  percent_records_cumulative	NUMBER
  lift_cumulative	NUMBER
  target_density_cumulative	NUMBER
  targets_cumulative	NUMBER
  non_targets_cumulative	NUMBER
  lift_quantile	NUMBER
  target_density	NUMBER

  
  

  See Also:

  Oracle Data Mining Concepts for details about the information in the lift table

• When a cost matrix is passed to COMPUTE_LIFT, the cost threshold is returned in the probability_threshold column of the lift table.

Examples

This example uses the Naive Bayes model nb_sh_clas_sample, which is created by one of the Oracle Data Mining sample programs.

The example illustrates lift based on probabilities. For examples that show computation based on costs, see "COMPUTE_CONFUSION_MATRIX Procedure".

Compute Lift Based on Probabilities

The following statement applies the model to the test data and stores the predictions and probabilities in a table.

CREATE TABLE nb_apply_results AS
          SELECT cust_id,
                      PREDICTION(nb_sh_clas_sample USING *) prediction,
                       PREDICTION_PROBABILITY(nb_sh_clas_sample USING *) probability
            FROM mining_data_test_v;

Using probabilities as the scoring criterion, you can compute lift as follows.

BEGIN
           DBMS_DATA_MINING.COMPUTE_LIFT (
              apply_result_table_name              => 'nb_apply_results',
              target_table_name                => 'mining_data_test_v',
              case_id_column_name              => 'cust_id',
              target_column_name               => 'affinity_card',
              lift_table_name                             => 'nb_lift',
              positive_target_value                 =>  to_char(1),
              score_column_name                => 'PREDICTION',
              score_criterion_column_name    => 'PROBABILITY',
              num_quantiles                                =>  10,
              cost_matrix_table_name                =>  null,
              apply_result_schema_name            =>  null,
              target_schema_name               =>  null,
              cost_matrix_schema_name              =>  null,
              score_criterion_type                   =>  'PROBABILITY');
        END;
        /

This query displays some of the statistics from the resulting lift table.

SQL>SELECT quantile_number, probability_threshold, gain_cumulative,
           quantile_total_count
           FROM nb_lift;

QUANTILE_NUMBER PROBABILITY_THRESHOLD GAIN_CUMULATIVE QUANTILE_TOTAL_COUNT 
--------------- --------------------- --------------- --------------------  
              1            .989335775       .15034965                   55 
              2            .980534911       .26048951                   55  
              3            .968506098      .374125874                   55  
              4            .958975196      .493006993                   55 
              5            .946705997      .587412587                   55  
              6            .927454174       .66958042                   55  
              7            .904403627      .748251748                   55  
              8            .836482525      .839160839                   55  
             10            .500184953               1                   54  

COMPUTE_ROC Procedure

This procedure computes receiver operating characteristic (ROC), stores the results in a table in the user's schema, and returns a measure of the model accuracy.

ROC is a test metric for binary classification models. To compute ROC, one of the target values must be designated as the positive class. COMPUTE_ROC compares the predictions generated by the model with the actual target values in a set of test data.

ROC measures the impact of changes in the probability threshold. The probability threshold is the decision point used by the model for predictions. In binary classification, the default probability threshold is 0.5. The value predicted for each case is the one with a probability greater than 50%.

ROC can be plotted as a curve on an X-Y axis. The false positive rate is placed on the X axis. The true positive rate is placed on the Y axis. A false positive is a positive prediction for a case that is negative in the test data. A true positive is a positive prediction for a case that is positive in the test data.

COMPUTE_ROC accepts two input streams:

• The predictions generated on the test data. The information is passed in three columns:

  • Case ID column

  • Prediction column

  • Scoring criterion column containing probabilities

• The known target values in the test data. The information is passed in two columns:

  • Case ID column

  • Target column containing the known target values

Syntax

DBMS_DATA_MINING.COMPUTE_ROC (
      roc_area_under_curve         OUT NUMBER,
      apply_result_table_name      IN  VARCHAR2,
      target_table_name            IN  VARCHAR2,
      case_id_column_name          IN  VARCHAR2,
      target_column_name           IN  VARCHAR2,
      roc_table_name               IN  VARCHAR2,
      positive_target_value        IN  VARCHAR2,
      score_column_name            IN  VARCHAR2 DEFAULT 'PREDICTION',
      score_criterion_column_name  IN  VARCHAR2 DEFAULT 'PROBABILITY',
      apply_result_schema_name     IN  VARCHAR2 DEFAULT NULL,
      target_schema_name           IN  VARCHAR2 DEFAULT NULL);

Parameters

Usage Notes

• The predictive information you pass to COMPUTE_ROC may be generated using SQL PREDICTION functions, the DBMS_DATA_MINING.APPLY procedure, or some other mechanism. As long as you pass the appropriate data, the procedure can compute the receiver operating characteristic.

• The predictions that you pass to COMPUTE_ROC are in a table or view specified in apply_results_table_name.

  CREATE TABLE apply_result_table_name AS (
              case_id_column_name            VARCHAR2, 
              score_column_name              VARCHAR2,
              score_criterion_column_name    VARCHAR2);
  

• The table created by COMPUTE_ROC has the columns shown in Table 42-35.

  Table 42-35 COMPUTE_ROC Output

  Column	Data Type
  probability	NUMBER
  true_positives	NUMBER
  false_negatives	NUMBER
  false_positives	NUMBER
  true_negatives	NUMBER
  true_positive_fraction	NUMBER
  false_positive_fraction	NUMBER

  
  

  See Also:

  Oracle Data Mining Concepts for details about the output of COMPUTE_ROC

• ROC is typically used to determine the most desirable probability threshold. This can be done by examining the true positive fraction and the false positive fraction. The true positive fraction is the percentage of all positive cases in the test data that were correctly predicted as positive. The false positive fraction is the percentage of all negative cases in the test data that were incorrectly predicted as positive.

  Given a probability threshold, the following statement returns the positive predictions in an apply result table ordered by probability.

  SELECT case_id_column_name 
         FROM apply_result_table_name 
         WHERE probability > probability_threshold 
         ORDER BY probability DESC;
  

• There are two approaches to identifying the most desirable probability threshold. Which approach you use depends on whether or not you know the relative cost of positive versus negative class prediction errors.

  If the costs are known, you can apply the relative costs to the ROC table to compute the minimum cost probability threshold. Suppose the relative cost ratio is: Positive Class Error Cost / Negative Class Error Cost = 20. Then execute a query like this.

  WITH cost AS (
    SELECT probability_threshold, 20 * false_negatives + false_positives cost 
      FROM ROC_table 
    GROUP BY probability_threshold), 
      minCost AS (
        SELECT min(cost) minCost 
          FROM cost)
        SELECT max(probability_threshold)probability_threshold 
          FROM cost, minCost 
      WHERE cost = minCost;
  

  If relative costs are not well known, you can simply scan the values in the ROC table (in sorted order) and make a determination about which of the displayed trade-offs (misclassified positives versus misclassified negatives) is most desirable.

  SELECT * FROM ROC_table 
           ORDER BY probability_threshold;
  

Examples

This example uses the Naive Bayes model nb_sh_clas_sample, which is created by one of the Oracle Data Mining sample programs.

The following statement applies the model to the test data and stores the predictions and probabilities in a table.

CREATE TABLE nb_apply_results AS
          SELECT cust_id,
                      PREDICTION(nb_sh_clas_sample USING *) prediction,
                       PREDICTION_PROBABILITY(nb_sh_clas_sample USING *) probability
            FROM mining_data_test_v;

Using the predictions and the target values from the test data, you can compute ROC as follows.

DECLARE
     v_area_under_curve NUMBER;
  BEGIN
         DBMS_DATA_MINING.COMPUTE_ROC (
               roc_area_under_curve                  => v_area_under_curve,
               apply_result_table_name       => 'nb_apply_results',
               target_table_name               => 'mining_data_test_v',
               case_id_column_name            => 'cust_id',
               target_column_name             => 'affinity_card',
               roc_table_name                     => 'nb_roc',
               positive_target_value         => '1',
               score_column_name               => 'PREDICTION',
               score_criterion_column_name   => 'PROBABILITY');
           DBMS_OUTPUT.PUT_LINE('**** AREA UNDER ROC CURVE ****: ' ||
           ROUND(v_area_under_curve,4));
  END;
 /

The resulting AUC and a selection of columns from the ROC table are shown as follows.

**** AREA UNDER ROC CURVE ****: .8212

SQL> SELECT probability, true_positive_fraction, false_positive_fraction 
            FROM nb_roc;
 
PROBABILITY  TRUE_POSITIVE_FRACTION  FALSE_POSITIVE_FRACTION
-----------  ----------------------  -----------------------
     .00000                       1                        1
     .50018              .826589595               .227902946
     .53851              .823699422               .221837088
     .54991              .820809249               .217504333
     .55628              .815028902               .215771231
     .55628              .817919075               .215771231
     .57563              .800578035               .214904679
     .57563              .812138728               .214904679
      .                   .                        .
      .                   .                        .
      .                   .                        .

CREATE_MODEL Procedure

This procedure creates a mining model with a given mining function.

By passing an xform_list to CREATE_MODEL, you can specify a list of transformations to be performed on the input data. If the PREP_AUTO setting is on, the transformations are used in addition to the automatic transformations. If the PREP_AUTO setting is off, the specified transformations are the only ones implemented by the model. In both cases, the transformation definitions are embedded in the model and executed automatically whenever the model is applied. See "Automatic Data Preparation".

Syntax

DBMS_DATA_MINING.CREATE_MODEL (
      model_name            IN VARCHAR2,
      mining_function       IN VARCHAR2,
      data_table_name       IN VARCHAR2,
      case_id_column_name   IN VARCHAR2,
      target_column_name    IN VARCHAR2 DEFAULT NULL,
      settings_table_name   IN VARCHAR2 DEFAULT NULL,
      data_schema_name      IN VARCHAR2 DEFAULT NULL,
      settings_schema_name  IN VARCHAR2 DEFAULT NULL,
      xform_list            IN TRANSFORM_LIST DEFAULT NULL);

Parameters

TYPE
  TRANFORM_REC     IS RECORD (
     attribute_name       VARCHAR2(4000),
     attribute_subname    VARCHAR2(4000),
     expression           EXPRESSION_REC,
     reverse_expression   EXPRESSION_REC,
     attribute_spec       VARCHAR2(4000));

Usage Notes

You can obtain information about a model by querying these data dictionary views.

Specify the USER prefix instead of ALL to obtain information about models in your own schema only.

Examples

The first example builds a classification model using the Support Vector Machine algorithm.

-- Create the settings table 
CREATE TABLE svm_model_settings (
  setting_name  VARCHAR2(30),
  setting_value VARCHAR2(30));

-- Populate the settings table
-- Specify SVM. By default, Naive Bayes is used for classification.
-- Specify ADP. By default, ADP is not used.
BEGIN 
  INSERT INTO svm_model_settings (setting_name, setting_value) VALUES
     (dbms_data_mining.algo_name, dbms_data_mining.algo_support_vector_machines);
  INSERT INTO svm_model_settings (setting_name, setting_value) VALUES
     (dbms_data_mining.prep_auto,dbms_data_mining.prep_auto_on);
  COMMIT;
END;
/
-- Create the model using the specified settings 
BEGIN
  DBMS_DATA_MINING.CREATE_MODEL(
    model_name          => 'svm_model',
    mining_function     => dbms_data_mining.classification,
    data_table_name     => 'mining_data_build_v',
    case_id_column_name => 'cust_id',
    target_column_name  => 'affinity_card',
    settings_table_name => 'svm_model_settings');
END;
/

You can display the model settings with the following query.

SELECT * FROM user_mining_model_settings 
       WHERE model_name IN 'SVM_MODEL';

MODEL_NAME     SETTING_NAME            SETTING_VALUE                  SETTING
-------------  ----------------------  -----------------------------  -------
SVM_MODEL      ALGO_NAME               ALGO_SUPPORT_VECTOR_MACHINES  INPUT
SVM_MODEL      SVMS_KERNEL_CACHE_SIZE  50000000                      DEFAULT
SVM_MODEL      SVMS_ACTIVE_LEARNING    SVMS_AL_ENABLE                DEFAULT
SVM_MODEL      SVMS_STD_DEV            3.004524                      DEFAULT
SVM_MODEL      PREP_AUTO               ON                            INPUT
SVM_MODEL      SVMS_COMPLEXITY_FACTOR  1.887389                      DEFAULT
SVM_MODEL      SVMS_KERNEL_FUNCTION    SVMS_GAUSSIAN                 DEFAULT
SVM_MODEL      SVMS_CONV_TOLERANCE     .001                          DEFAULT

The second example creates an anomaly detection model. Anomaly detection uses SVM classification without a target. This example uses the same settings table created for the SVM classification model in the first example.

BEGIN
  DBMS_DATA_MINING.CREATE_MODEL(
    model_name          => 'anomaly_detect_model',
    mining_function     => dbms_data_mining.classification,
    data_table_name     => 'mining_data_build_v',
    case_id_column_name => 'cust_id',
    target_column_name  => null,
    settings_table_name => 'svm_model_settings');
END;
/

This query shows that the models created in these examples are the only ones in your schema.

SELECT model_name, mining_function, algorithm FROM user_mining_models;
 
MODEL_NAME              MINING_FUNCTION      ALGORITHM
----------------------  -------------------- ------------------------------
SVM_MODEL               CLASSIFICATION       SUPPORT_VECTOR_MACHINES
ANOMALY_DETECT_MODEL    CLASSIFICATION       SUPPORT_VECTOR_MACHINES

This query shows that only the SVM classification model has a target.

SELECT model_name, attribute_name, attribute_type, target 
       FROM user_mining_model_attributes 
       WHERE target = 'YES';
 
MODEL_NAME          ATTRIBUTE_NAME   ATTRIBUTE_TYPE     TARGET
------------------  ---------------  -----------------  ------
SVM_MODEL           AFFINITY_CARD    CATEGORICAL         YES

DROP_MODEL Procedure

This procedure deletes the specified mining model.

Syntax

DBMS_DATA_MINING.DROP_MODEL (model_name IN VARCHAR2,
                             force      IN BOOLEAN DEFAULT FALSE);

Parameters

Usage Note

To drop a mining model, you must be the owner or you must have the DROP ANY MINING MODEL privilege. See Oracle Data Mining Administrator's Guide for information about privileges for data mining.

Example

You can use the following command to delete a valid mining model named nb_sh_clas_sample that exists in your schema.

BEGIN
  DBMS_DATA_MINING.DROP_MODEL(model_name => 'nb_sh_clas_sample');
END;
/

EXPORT_MODEL Procedure

This procedure exports the specified data mining models to a dump file set. You can import from the dump file set using the IMPORT_MODEL procedure. Both EXPORT_MODEL and IMPORT_MODEL use Oracle Data Pump technology.

Syntax

DBMS_DATA_MINING.EXPORT_MODEL (
      filename          IN VARCHAR2,
      directory         IN VARCHAR2,
      model_filter      IN VARCHAR2 DEFAULT NULL,
      filesize          IN VARCHAR2 DEFAULT NULL,
      operation         IN VARCHAR2 DEFAULT NULL,
      remote_link       IN VARCHAR2 DEFAULT NULL,
      jobname           IN VARCHAR2 DEFAULT NULL);

Parameters

Usage Notes

The model_filter parameter specifies which models to export. You can list the models by name, or you can specify all models that have the same mining function or algorithm. You can query the USER_MINING_MODELS view to list the models in your schema.

SQL> describe user_mining_models
 Name                                      Null?    Type
 ----------------------------------------- -------- ----------------------------
 MODEL_NAME                                NOT NULL VARCHAR2(30)
 MINING_FUNCTION                                    VARCHAR2(30)
 ALGORITHM                                          VARCHAR2(30)
 CREATION_DATE                             NOT NULL DATE
 BUILD_DURATION                                     NUMBER
 MODEL_SIZE                                         NUMBER
 COMMENTS                                           VARCHAR2(4000)

For more information on data dictionary views of mining models, see "Mining Model Objects".

Examples of model filters are provided in Table 42-39.

Examples

The following statement exports all the models in the DMUSER3 schema to a dump file set called models_out in the directory $ORACLE_HOME/rdbms/log. This directory is mapped to a directory object called DATA_PUMP_DIR. The DMUSER3 user has read/write access to the directory and to the directory object.

SQL>execute dbms_data_mining.export_model ('models_out', 'DATA_PUMP_DIR');

You can exit SQL*Plus and list the resulting dump file and log file.

SQL>exit
>cd $ORACLE_HOME/rdbms/log
>ls
>DMUSER3_exp_1027.log  models_out01.dmp  

The following example uses the same directory object and is executed by the same user. It exports the models called NMF_SH_SAMPLE and SVMR_SH_REGR_SAMPLE to a different dump file set in the same directory.

SQL>execute dbms_data_mining.export_model ( 'models2_out', 'DATA_PUMP_DIR',
            'name in (''NMF_SH_SAMPLE'', ''SVMR_SH_REGR_SAMPLE'')');
SQL>exit
>cd $ORACLE_HOME/rdbms/log
>ls
>DMUSER3_exp_1027.log  models_out01.dmp
 DMUSER3_exp_924.log  models2_out01.dmp

GET_ASSOCIATION_RULES Function

This table function returns the rules from an association model.

You can specify filtering criteria to cause GET_ASSOCIATION_RULES to return a subset of the rules. Filtering criteria can improve the performance of the table function. If the number of rules is large, the greatest performance improvement will result from specifying the topn parameter.

Syntax

DBMS_DATA_MINING.GET_ASSOCIATION_RULES (
   model_name            IN VARCHAR2,
   topn                  IN NUMBER DEFAULT NULL,
   rule_id               IN INTEGER DEFAULT NULL,
   min_confidence        IN NUMBER DEFAULT NULL,
   min_support           IN NUMBER DEFAULT NULL,
   max_rule_length       IN INTEGER DEFAULT NULL,
   min_rule_length       IN INTEGER DEFAULT NULL,
   sort_order            IN ORA_MINING_VARCHAR2_NT DEFAULT NULL,
   antecedent_items      IN DM_ITEMS DEFAULT NULL,
   consequent_items      IN DM_ITEMS DEFAULT NULL,
   min_lift              IN NUMBER DEFAULT NULL)
 RETURN DM_RULES PIPELINED;

Parameters

Return Values

(rule_id              INTEGER,
 antecedent           DM_PREDICATES,
 consequent           DM_PREDICATES,
 rule_support         NUMBER,
 rule_confidence      NUMBER,
 rule_lift            NUMBER,
 antecedent_support   NUMBER,
 consequent_support   NUMBER,
 number_of_items      INTEGER )

(attribute_name            VARCHAR2(4000),
      attribute_subname         VARCHAR2(4000),
      conditional_operator      CHAR(2)/*=,<>,<,>,<=,>=*/,
      attribute_num_value       NUMBER,
      attribute_str_value       VARCHAR2(4000),
      attribute_support         NUMBER,
      attribute_confidence      NUMBER)

Usage Notes

This table function pipes out rows of type DM_RULES. For information on Data Mining data types and piped output from table functions, see "Data Types".

The ORA_MINING_VARCHAR2_NT type is defined as a table of VARCHAR2(4000).

Examples

The following example demonstrates an Association model build followed by several invocations of the GET_ASSOCIATION_RULES table function.

-- prepare a settings table to override default settings
CREATE TABLE market_settings AS
SELECT *
  FROM TABLE(DBMS_DATA_MINING.GET_DEFAULT_SETTINGS)
 WHERE setting_name LIKE 'ASSO_%';
BEGIN
-- update the value of the minimum confidence
UPDATE census_settings
   SET setting_value = TO_CHAR(0.081)
 WHERE setting_name = DBMS_DATA_MINING.asso_min_confidence;

-- build an AR model 
DBMS_DATA_MINING.CREATE_MODEL(
  model_name => 'market_model',
  function => DBMS_DATA_MINING.ASSOCIATION,
  data_table_name => 'market_build',
  case_id_column_name => 'item_id',
  target_column_name => NULL,
  settings_table_name => 'census_settings');
END;
/
-- View the (unformatted) rules 
SELECT rule_id, antecedent, consequent, rule_support,
       rule_confidence
  FROM TABLE(DBMS_DATA_MINING.GET_ASSOCIATION_RULES('market_model'));

In the previous example, you view all rules. To view just the top 20 rules, use the following statement.

-- View the top 20 (unformatted) rules
SELECT rule_id, antecedent, consequent, rule_support,
       rule_confidence
  FROM TABLE(DBMS_DATA_MINING.GET_ASSOCIATION_RULES('market_model', 20));

The following example returns all the rules which have 'AQUATIC' or 'EGGS' in the antecedent, and has 'VENOMOUS' as the consequent. The rules are sorted first by NUMBER_OF_ITEMS in descending order, then by RULE_CONFIDENCE in descending order, and finally by RULE_SUPPORT in descending order.

SELECT * FROM TABLE
(    DBMS_DATA_MINING.GET_ASSOCIATION_RULES
       ('AR_Model_31', 120, NULL, 1, .51, 7,
         ORA_MINING_VARCHAR2_NT
          ('NUMBER_OF_ITEMS DESC', 'RULE_CONFIDENCE DESC', 'RULE_SUPPORT DESC'),
         ORA_MINING_VARCHAR2_NT('AQUATIC', 'EGGS'),
         ORA_MINING_VARCHAR2_NT('VENOMOUS')));

GET_DEFAULT_SETTINGS Function

The GET_DEFAULT_SETTINGS function is deprecated in 11g Release 2 (11.2). It is replaced with the data dictionary view, *_MINING_MODEL_SETTINGS. USER_, ALL_, and DBA_ versions of the view are available. See Oracle Data Mining Application Developer's Guide.

Oracle recommends that you do not use deprecated procedures in new applications. Support for deprecated features is for backward compatibility only.

This table function returns the default settings for all mining functions and algorithms supported in the DBMS_DATA_MINING package.

Syntax

DBMS_DATA_MINING.GET_DEFAULT_SETTINGS
  RETURN DM_MODEL_SETTINGS PIPELINED;

Return Values

(setting_name    VARCHAR2(30),
 setting_value   VARCHAR2(128))

Usage Notes

This table function pipes out rows of type DM_MODEL_SETTING. For information on Data Mining data types and piped output from table functions, see "Data Types".

This function is particularly useful if you do not know what settings are associated with a particular function or algorithm, and you want to override some or all of them.

Examples

For example, if you want to override some or all of k-Means clustering settings, you can create a settings table as shown, and update individual settings as required.

BEGIN
  CREATE TABLE mysettings AS
  SELECT * 
  FROM TABLE(DBMS_DATA_MINING.GET_DEFAULT_SETTINGS)
   WHERE setting_name LIKE 'KMNS%';
  -- now update individual settings as required
  UPDATE mysettings
     SET setting_value = 0.02
   WHERE setting_name = DBMS_DATA_MINING.KMNS_MIN_PCT_ATTR_SUPPORT;
END;
/

GET_FREQUENT_ITEMSETS Function

This table function returns a set of rows that represent the frequent itemsets from an Association model. For a detailed description of frequent itemsets, consult Oracle Data Mining Concepts.

Syntax

DBMS_DATA_MINING.GET_FREQUENT_ITEMSETS (
    model_name          IN VARCHAR2,
    topn                IN NUMBER DEFAULT NULL,
    max_itemset_length  IN NUMBER DEFAULT NULL)
  RETURN DM_ITEMSETS PIPELINED;

Parameters

Return Values

(itemsets_id      NUMBER,
items             DM_ITEMS,
support           NUMBER,
number_of_items   NUMBER)

(attribute_name      VARCHAR2(4000),
attribute_subname    VARCHAR2(4000),
attribute_num_value  NUMBER,
attribute_str_value  VARCHAR2(4000))

Usage Notes

This table function pipes out rows of type DM_ITEMSETS. For information on Data Mining data types and piped output from table functions, see "Data Types".

Examples

The following example demonstrates an Association model build followed by an invocation of GET_FREQUENT_ITEMSETS table function from Oracle SQL.

-- prepare a settings table to override default settings
CREATE TABLE market_settings AS

SELECT *

FROM TABLE(DBMS_DATA_MINING.GET_DEFAULT_SETTINGS)
 WHERE setting_name LIKE 'ASSO_%';
BEGIN
-- update the value of the minimum confidence
UPDATE market_settings
   SET setting_value = TO_CHAR(0.081)
 WHERE setting_name = DBMS_DATA_MINING.asso_min_confidence;

/* build a AR model */
DBMS_DATA_MINING.CREATE_MODEL(
  model_name           => 'market_model',
  function             => DBMS_DATA_MINING.ASSOCIATION,
  data_table_name      => 'market_build',
  case_id_column_name  => 'item_id',
  target_column_name   => NULL,
  settings_table_name  => 'census_settings');
END;
/

-- View the (unformatted) Itemsets from SQL*Plus
SELECT itemset_id, items, support, number_of_items
  FROM TABLE(DBMS_DATA_MINING.GET_FREQUENT_ITEMSETS('market_model'));

In the example above, you view all itemsets. To view just the top 20 itemsets, use the following statement:

-- View the top 20 (unformatted) Itemsets from SQL*Plus
SELECT itemset_id, items, support, number_of_items
  FROM TABLE(DBMS_DATA_MINING.GET_FREQUENT_ITEMSETS('market_model', 20));

GET_MODEL_COST_MATRIX Function

This function returns the rows of the default scoring matrix associated with the specified model.

By default, this function returns the scoring matrix that was added to the model with the ADD_COST_MATRIX procedure. If you wish to obtain the cost matrix used to create a model, specify cost_matrix_type_create as the matrix_type. See Table 42-45.

See also ADD_COST_MATRIX Procedure.

Syntax

DBMS_DATA_MINING.GET_MODEL_COST_MATRIX (
      model_name        IN VARCHAR2,
      matrix_type       IN VARCHAR2 DEFAULT cost_matrix_type_score)
RETURN DM_COST_MATRIX PIPELINED;

Parameters

Return Values

actual          VARCHAR2(4000), predicted       VARCHAR2(4000), cost            NUMBER)

Usage Notes

Only Decision Tree models can be built with a cost matrix. If you want to build a Decision Tree model with a cost matrix, specify the cost matrix table name in the CLAS_COST_TABLE_NAME setting in the settings table for the model. See Table 42-7, "Mining Function Settings".

The cost matrix used to create a Decision Tree model becomes the default scoring matrix for the model. If you want to specify different costs for scoring, you can modify the values in the cost matrix table or you can use the REMOVE_COST_MATRIX procedure to remove the cost matrix and the ADD_COST_MATRIX procedure to add a new one

Example

This example returns the scoring cost matrix associated with the Naive Bayes model NB_SH_CLAS_SAMPLE.

column actual format a10
column predicted format a10
SELECT *
    FROM TABLE(dbms_data_mining.get_model_cost_matrix('nb_sh_clas_sample'))
    ORDER BY predicted, actual;
 
ACTUAL     PREDICTED   COST
---------- ---------- -----
0          0            .00
1          0            .75
0          1            .25
1          1            .00

GET_MODEL_DETAILS_ABN Function

The Adaptive Bayes Network algorithm ABN algorithm is deprecated in 11g Release 2 (11.2).

Oracle recommends that you do not use deprecated procedures in new applications. Support for deprecated features is for backward compatibility only.

This table function returns a set of rows that provide the details of an Adaptive Bayes Network model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_ABN (
    model_name         IN VARCHAR2)
  RETURN DM_ABN_DETAILS PIPELINED;

Parameters

Return Values

(rule_id           INTEGER,
 antecedent        DM_PREDICATES,
 consequent        DM_PREDICATES,
 rule_support      NUMBER)

(attribute_name          VARCHAR2(4000),
      attribute_subname       VARCHAR2(4000),
      conditional_operator    CHAR(2), /*=,<>,<,>,<=,>=*/
      attribute_num_value     NUMBER,
      attribute_str_value     VARCHAR2(4000),
      attribute_support       NUMBER,
      attribute_confidence    NUMBER)

Usage Notes

This table function pipes out rows of type DM_ABN_DETAIL. For information on Data Mining data types and piped output from table functions, see "Data Types".

This function returns details only for a single feature ABN model.

Examples

The following example demonstrates an ABN model build followed by an invocation of GET_MODEL_DETAILS_ABN table function from Oracle SQL.

BEGIN
  -- prepare a settings table to override default algorithm and model type
  CREATE TABLE abn_settings (setting_name VARCHAR2(30),
  setting_value 
VARCHAR2(128));
  INSERT INTO abn_settings VALUES (DBMS_DATA_MINING.ALGO_NAME,
    DBMS_DATA_MINING.ALGO_ADAPTIVE_BAYES_NETWORK);
  INSERT INTO abn_settings VALUES    (DBMS_DATA_MINING.ABNS_MODEL_TYPE,     DBMS_DATA_MINING.ABNS_SINGLE_FEATURE);
   COMMIT;
  -- create a model
  DBMS_DATA_MINING.CREATE_MODEL (
    model_name           => 'abn_model',
    function             => DBMS_DATA_MINING.CLASSIFICATION,
    data_table_name      => 'abn_build',
    case_id_column_name  => 'id',
    target_column_name   => NULL,
    settings_table_name  => 'abn_settings');
END;
/
-- View the (unformatted) results from SQL*Plus
SELECT *
    FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_ABN('abn_model'));

GET_MODEL_DETAILS_AI Function

This table function returns a set of rows that provide the details of an Attribute Importance model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_AI (
  model_name         IN VARCHAR2)
 RETURN DM_RANKED_ATTRIBUTES PIPELINED;

Parameters

Return Values

(attribute_name          VARCHAR2(4000,
 attribute_subname       VARCHAR2(4000),
 importance_value        NUMBER,
 rank                    NUMBER(38))

GET_MODEL_DETAILS_GLM Function

This table function returns the coefficient statistics for a Generalized Linear Model.

The same set of statistics is returned for both linear and logistic regression, but statistics that do not apply to the mining function are returned as NULL. For more details, see the Usage Notes.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_GLM (
             model_name             VARCHAR2)
RETURN DM_GLM_COEFF_SET PIPELINED;

Parameters

Return Values

(class                   VARCHAR2(4000),
 attribute_name          VARCHAR2(4000),
 attribute_subname       VARCHAR2(4000),
 attribute_value         VARCHAR2(4000),
 coefficient             NUMBER,
 std_error               NUMBER,
 test_statistic          NUMBER,
 p_value                 NUMBER,
 VIF                     NUMBER,
 std_coefficient         NUMBER,
 lower_coeff_limit       NUMBER,
 upper_coeff_limit       NUMBER,
 exp_coefficient         BINARY_DOUBLE,
 exp_lower_coeff_limit   BINARY_DOUBLE,
 exp_upper_coeff_limit   BINARY_DOUBLE)

GET_MODEL_DETAILS_GLM returns a row of statistics for each attribute and one extra row for the intercept, which is identified by a null value in the attribute name. Each row has the DM_GLM_COEFF data type. The statistics are described in Table 42-53.

Usage Notes

Not all statistics are necessarily returned for each coefficient. Statistics will be null if:

• They do not apply to the mining function. For example, exp_coefficient does not apply to linear regression.

• They cannot be computed from a theoretical standpoint. For example, when ridge regression is enabled, the coefficient values are returned with no statistics except VIF if it is enabled. (For information on ridge regression, see Table 42-11, "GLM Settings".)

• They cannot be computed because of limitations in system resources.

• Their values would be infinity.

GET_MODEL_DETAILS_GLOBAL Function

This table function returns statistics about the model as a whole. Global details are available for GLM and for association rules.

Separate global details are returned for linear and logistic regression. When ridge regression is enabled, fewer global details are returned. For information about ridge, see Table 42-11, "GLM Settings".

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_GLOBAL (
      model_name     IN  VARCHAR2)
RETURN DM_MODEL_GLOBAL_DETAILS PIPELINED;

Parameters

Return Values

(global_detail_name   VARCHAR2(30),
 global_detail_value   NUMBER)

Global Details for GLM: Linear Regression

Global Details for GLM: Logistic Regression

Global Detail for Association Rules

A single global detail is produced by an Association model.

GET_MODEL_DETAILS_KM Function

This table function returns a set of rows that provide the details of a k-Means clustering model.

You can provide input to GET_MODEL_DETAILS_KM to request specific information about the model, thus improving the performance of the query. If you do not specify filtering parameters, GET_MODEL_DETAILS_KM returns all the information about the model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_KM (
          model_name          VARCHAR2,
          cluster_id          NUMBER    DEFAULT NULL,
          attribute           VARCHAR2  DEFAULT NULL,
          centroid            NUMBER    DEFAULT 1, 
          histogram           NUMBER    DEFAULT 1, 
          rules               NUMBER    DEFAULT 2,
          attribute_subname   VARCHAR2  DEFAULT NULL)
RETURN DM_CLUSTERS PIPELINED;

Parameters

Return Values

(id                   NUMBER,
 cluster_id           VARCHAR2(4000),
 record_count         NUMBER,
 parent               NUMBER,
 tree_level           NUMBER,
 dispersion           NUMBER,
 split_predicate      DM_PREDICATES,
 child                DM_CHILDREN,
 centroid             DM_CENTROIDS,
 histogram            DM_HISTOGRAMS,
 rule                 DM_RULE)

(attribute_name           VARCHAR2(4000),
      attribute_subname        VARCHAR2(4000),
      conditional_operator     CHAR(2) /*=,<>,<,>,<=,>=*/,
      attribute_num_value      NUMBER,
      attribute_str_value      VARCHAR2(4000),
      attribute_support        NUMBER,
      attribute_confidence     NUMBER)

(attribute_name    VARCHAR2(4000),
      attribute_subname  VARCHAR2(4000),
      mean               NUMBER,
      mode_value         VARCHAR2(4000),
      variance           NUMBER)

(attribute_name    VARCHAR2(4000),
      attribute_subname  VARCHAR2(4000),
      bin_id             NUMBER,
      lower_bound        NUMBER,
      upper_bound        NUMBER,
      label              VARCHAR2(4000),
      count              NUMBER)

(rule_id            INTEGER,
      antecedent         DM_PREDICATES,
      consequent         DM_PREDICATES,
      rule_support       NUMBER,
      rule_confidence    NUMBER,
      rule_lift          NUMBER,
      antecedent_support NUMBER,
      consequent_support NUMBER,
      number_of_items    INTEGER)

(attribute_name           VARCHAR2(4000),
           attribute_subname  VARCHAR2(4000),
           conditional_operator     CHAR(2)/*=,<>,<,>,<=,>=*/,
           attribute_num_value      NUMBER,
           attribute_str_value      VARCHAR2(4000),
           attribute_support        NUMBER,
           attribute_confidence     NUMBER)

Usage Notes

The table function pipes out rows of type DM_CLUSTERS. For information on Data Mining data types and piped output from table functions, see "Data Types".

Examples

The following example demonstrates a k-Means clustering model build followed by an invocation of GET_MODEL_DETAILS_KM table function from Oracle SQL.

BEGIN
-- create a settings table
UPDATE cluster_settings
   SET setting_value = 3 
 WHERE setting_name = DBMS_DATA_MINING.KMEANS_BLOCK_GROWTH;

/* build a k-Means clustering model */
DBMS_DATA_MINING.CREATE_MODEL(
  model_name           => 'eight_clouds',
  function             => DBMS_DATA_MINING.CLUSTERING,
  data_table_name      => 'eight_clouds_build',
  case_id_column_name  => 'id',
  target_column_name   => NULL,
  settings_table_name  => 'cluster_settings');
END;
/

-- View the (unformatted) rules from SQL*Plus
SELECT id, record_count, parent, tree_level, dispersion,
       child, centroid, histogram, rule
  FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_KM('eight_clouds'));

GET_MODEL_DETAILS_NB Function

This table function returns a set of rows that provide the details of a Naive Bayes model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_NB (
   model_name      IN       VARCHAR2)
 RETURN DM_NB_DETAILS PIPELINED;

Parameters

Return Values

(target_attribute_name          VARCHAR2(30),
 target_attribute_str_value     VARCHAR2(4000),
 target_attribute_num_value     NUMBER,
 prior_probability              NUMBER,
 conditionals                   DM_CONDITIONALS)

(attribute_name             VARCHAR2(4000),
      attribute_subname          VARCHAR2(4000),
      attribute_str_value        VARCHAR2(4000),
      attribute_num_value        NUMBER,
      conditional_probability    NUMBER)

Usage Notes

The table function pipes out rows of type DM_NB_DETAILS. For information on Data Mining data types and piped output from table functions, see "Data Types".

Examples

Assume that you have built a classification model census_model using the Naive Bayes algorithm. You can retrieve the model details as shown in this example.

-- You can view the Naive Bayes model details in many ways
-- Consult the Oracle Application Developer's Guide -
-- Object-Relational Features for different ways of 
-- accessing Oracle Objects.

-- View the (unformatted) details from SQL*Plus
SELECT attribute_name, attribute_num_value, attribute_str_value,
       prior_probability, conditionals,
  FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_NB('census_model');

See nbdemo.sql for generation of formatted rules.

GET_MODEL_DETAILS_NMF Function

This table function returns a set of rows that provide the details of a Non-Negative Matrix Factorization model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_NMF (
   model_name        IN        VARCHAR2)
 RETURN DM_NMF_FEATURE_SET PIPELINED;

Parameters

Return Values

(feature_id          NUMBER,
 mapped_feature_id   VARCHAR2(4000),
 attribute_set       DM_NMF_ATTRIBUTE_SET)

(attribute_name    VARCHAR2(4000),
      attribute_subname  VARCHAR2(4000),
      attribute_value    VARCHAR2(4000),
      coefficient        NUMBER)

Usage Notes

The table function pipes out rows of type DM_NMF_FEATURE_SET. For information on Data Mining data types and piped output from table functions, see "Data Types".

Examples

Assume you have built an NMF model called my_nmf_model. You can retrieve model details as shown:

--View (unformatted) details from SQL*Plus 
SELECT feature_id, attribute_set 
FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_NMF(
        'my_nmf_model'));

GET_MODEL_DETAILS_OC Function

This table function returns a set of rows that provide the details of an O-Cluster clustering model. The rows are an enumeration of the clustering patterns generated during the creation of the model.

You can provide input to GET_MODEL_DETAILS_OC to request specific information about the model, thus improving the performance of the query. If you do not specify filtering parameters, GET_MODEL_DETAILS_OC returns all the information about the model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_OC (
          model_name         VARCHAR2,
          cluster_id         NUMBER    DEFAULT NULL,
          attribute          VARCHAR2  DEFAULT NULL,
          centroid           NUMBER    DEFAULT 1, 
          histogram          NUMBER    DEFAULT 1, 
          rules              NUMBER    DEFAULT 2)
RETURN DM_CLUSTERS PIPELINED;

Parameters

Return Values

(id               NUMBER,
 cluster_id       VARCHAR2(4000),
 record_count     NUMBER,
 parent           NUMBER,
 tree_level       NUMBER,
 dispersion       NUMBER,
 split_predicate  DM_PREDICATES,
 child            DM_CHILDREN,
 centroid         DM_CENTROIDS,
 histogram        DM_HISTOGRAMS,
 rule             DM_RULE)

(attribute_name           VARCHAR2(4000),
      attribute_subname        VARCHAR2(4000),
      conditional_operator     CHAR(2) /*=,<>,<,>,<=,>=*/,
      attribute_num_value      NUMBER,
      attribute_str_value      VARCHAR2(4000),
      attribute_support        NUMBER,
      attribute_confidence     NUMBER)

(attribute_name    VARCHAR2(4000),
       attribute_subname  VARCHAR2(4000),
       mean               NUMBER,
       mode_value         VARCHAR2(4000),
       variance           NUMBER)

(attribute_name    VARCHAR2(4000),
     attribute_subname  VARCHAR2(4000),
     bin_id             NUMBER,
     lower_bound        NUMBER,
     upper_bound        NUMBER,
     label              VARCHAR2(4000),
     count              NUMBER)

(rule_id            INTEGER,
      antecedent         DM_PREDICATES,
      consequent         DM_PREDICATES,
      rule_support       NUMBER,
      rule_confidence    NUMBER,
      rule_lift          NUMBER,
      antecedent_support NUMBER,
      consequent_support NUMBER,
      number_of_items    INTEGER)

(attribute_name           VARCHAR2(4000),
           attribute_subname        VARCHAR2(4000),
           conditional_operator     CHAR(2)/*=,<>,<,>,<=,>=*/,
           attribute_num_value      NUMBER,
           attribute_str_value      VARCHAR2(4000),
           attribute_support        NUMBER,
           attribute_confidence     NUMBER)

Usage Notes

The table function pipes out rows of type DM_CLUSTER. For information about Data Mining data types and piped output from table functions, see "Data Types".

Examples

Assume you have built an OC model called my_oc_model. You can retrieve information from the model details as shown:

--View (unformatted) details from SQL*Plus 
SELECT T.id           clu_id,
       T.record_count rec_cnt,
       T.parent       parent,
       T.tree_level   tree_level
  FROM (SELECT *
          FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_OC(
                 'my_oc_model'))
        ORDER BY id) T
 WHERE ROWNUM < 11;

GET_MODEL_DETAILS_SVM Function

This table function returns a set of rows that provide the details of a linear Support Vector Machine (SVM) model. If invoked for nonlinear SVM, it returns ORA-40215.

In linear SVM models, only nonzero coefficients are stored. This reduces storage and speeds up model loading. As a result, if an attribute is missing in the coefficient list returned by GET_MODEL_DETAILS_SVM, then the coefficient of this attribute should be interpreted as zero.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_SVM (
        model_name        VARCHAR2,
        reverse_coef      NUMBER DEFAULT 0)
 RETURN DM_SVM_LINEAR_COEFF_SET PIPELINED;

Parameters

Return Values

(class            VARCHAR2(4000),
 attribute_set    DM_SVM_ATTRIBUTE_SET)

(attribute_name      VARCHAR2(4000),
      attribute_subname   VARCHAR2(4000),
      attribute_value     VARCHAR2(4000),
      coefficient         NUMBER)

Usage Notes

1. This table function pipes out rows of type DM_SVM_LINEAR_COEFF. For information on Data Mining data types and piped output from table functions, see "Data Types".

2. The class column of DM_SVM_LINEAR_COEFF contains classification target values. For SVM regression models, class is null. For each classification target value, a set of coefficients is returned. For binary classification, one-class classifion, and regression models, only a single set of coefficients is returned.

3. The attribute_value column in DM_SVM_ATTRIBUTE_SET is used for categorical attributes.

4. GET_MODEL_DETAILS functions preserve model transparency by automatically reversing the transformations applied during the build process. Thus the attributes returned in the model details are the original attributes (or a close approximation of the original attributes) used to build the model.

   The coefficients are related to the transformed, not the original, attributes. When returned directly with the model details, the coefficients may not provide meaningful information. If you want GET_MODEL_DETAILS_SVM to transform the coefficients such that they relate to the original attributes, set the reverse_coef parameter to 1.

Examples

The following example demonstrates an SVM model build followed by an invocation of GET_MODEL_DETAILS_SVM table function from Oracle SQL:

-- Create SVM model
BEGIN
  dbms_data_mining.create_model(
    model_name           => 'SVM_Clas_sample',
    mining_function      => dbms_data_mining.classification,
    data_table_name      => 'svmc_sample_build_prepared',
    case_id_column_name  => 'id',
    target_column_name   => 'affinity_card',
    settings_table_name  => 'svmc_sample_settings');
END;
/
-- Display model details
SELECT *
  FROM TABLE(DBMS_DATA_MINING.GET_MODEL_DETAILS_SVM('SVM_Clas_sample'))
ORDER BY class;

GET_MODEL_DETAILS_XML Function

This function returns an XML object that provides the details of a Decision Tree model.

Syntax

DBMS_DATA_MINING.GET_MODEL_DETAILS_XML (
         model_name      IN       VARCHAR2)
   RETURN XMLTYPE;

Parameters

Return Values

Usage Notes

Special characters that cannot be displayed by Oracle XML are converted to '#'.

Examples

The following statements in SQL*Plus return the details of the decision tree model dt_sh_clas_sample. This model is created by the program dmdtdemo.sql, one of the sample data mining programs provided with Oracle Database Examples.

Note: The "&quot" characters you will see in the XML output are a result of SQL*Plus behavior. To display the XML in proper format, cut and past it into a file and open the file in a browser.

column dt_details format a320
EXEC dbms_xdb_print.setPrintMode(dbms_xdb_print.PRINT_PRETTY, 2);
SELECT 
 dbms_data_mining.get_model_details_xml('dt_sh_clas_sample') 
 AS DT_DETAILS
FROM dual;



DT_DETAILS
--------------------------------------------------------------------------------
<PMML version="2.1">
  <Header copyright="Copyright (c) 2004, Oracle Corporation. All rights
      reserved."/>
  <DataDictionary numberOfFields="9">
    <DataField name="AFFINITY_CARD" optype="categorical"/> 
    <DataField name="AGE" optype="continuous"/> 
    <DataField name="BOOKKEEPING_APPLICATION" optype="continuous"/>
    <DataField name="CUST_MARITAL_STATUS" optype="categorical"/>
    <DataField name="EDUCATION" optype="categorical"/> 
    <DataField name="HOUSEHOLD_SIZE" optype="categorical"/>
    <DataField name="OCCUPATION" optype="categorical"/>
    <DataField name="YRS_RESIDENCE" optype="continuous"/>
    <DataField name="Y_BOX_GAMES" optype="continuous"/>
  </DataDictionary>
  <TreeModel modelName="DT_SH_CLAS_SAMPLE" functionName="classification"
      splitCharacteristic="binarySplit">
    <Extension name="buildSettings">
      <Setting name="TREE_IMPURITY_METRIC" value="TREE_IMPURITY_GINI"/>
      <Setting name="TREE_TERM_MAX_DEPTH" value="7"/>
      <Setting name="TREE_TERM_MINPCT_NODE" value=".05"/>
      <Setting name="TREE_TERM_MINPCT_SPLIT" value=".1"/> 
      <Setting name="TREE_TERM_MINREC_NODE" value="10"/>
      <Setting name="TREE_TERM_MINREC_SPLIT" value="20"/>
      <costMatrix>
        <costElement>
          <actualValue>0</actualValue>
          <predictedValue>0</predictedValue>
          <cost>0</cost>
        </costElement>
        <costElement>
          <actualValue>0</actualValue>
          <predictedValue>1</predictedValue>
          <cost>1</cost>
        </costElement>
        <costElement>
          <actualValue>1</actualValue>
          <predictedValue>0</predictedValue>
          <cost>8</cost> 
        </costElement>
        <costElement> 
          <actualValue>1</actualValue>
          <predictedValue>1</predictedValue> 
          <cost>0</cost> 
        </costElement>
      </costMatrix>
    </Extension>
    <MiningSchema>
      .
      .
      .
      .
      .
      . 
      </Node>
    </Node>
  </TreeModel>
</PMML> 

GET_MODEL_SETTINGS Function

The GET_MODEL_SETTINGS function is deprecated in 11g Release 2 (11.2). It is replaced with the data dictionary view, *_MINING_MODEL_SETTINGS. USER_, ALL_, and DBA_ versions of the view are available. See Oracle Data Mining Application Developer's Guide.

Oracle recommends that you do not use deprecated procedures in new applications. Support for deprecated features is for backward compatibility only.

This table function returns the list of settings that were used to build the model.

Syntax

DBMS_DATA_MINING.GET_MODEL_SETTINGS(
   model_name           IN VARCHAR2)
 RETURN DM_MODEL_SETTINGS PIPELINED;

Parameters

Return Values

(setting_name    VARCHAR2(30),
setting_value    VARCHAR2(128))

Usage Notes

The table function pipes out rows of type DM_MODEL_SETTING. For information about Data Mining data types and piped output from table functions, see "Data Types".

You can use this table function to determine the settings that were used to build the model. This is purely for informational purposes only — you cannot alter the model to adopt new settings.

Examples

Assume that you have built a classification model census_model using the Naive Bayes algorithm. You can retrieve the model settings using Oracle SQL as follows:

SELECT setting_name, setting_value
  FROM TABLE(DBMS_DATA_MINING.GET_MODEL_SETTINGS('census_model'));

GET_MODEL_SIGNATURE Function

The GET_MODEL_SIGNATURE function is deprecated in 11g Release 2 (11.2). It is replaced with the data dictionary view, *_MINING_MODEL_ATTRIBUTES. USER_, ALL_, and DBA_ versions of the view are available. See Oracle Data Mining Application Developer's Guide.

Oracle recommends that you do not use deprecated procedures in new applications. Support for deprecated features is for backward compatibility only.

This table function returns the model signature, which lists the column attributes used to build the model and which should be present in the scoring data.

The case identifier is not considered a mining attribute. For classification and regression models, the target attribute is also not considered part of the model signature.

Syntax

DBMS_DATA_MINING.GET_MODEL_SIGNATURE(
  model_name           IN VARCHAR2)
RETURN DM_MODEL_SIGNATURE PIPELINED;

Parameters

Return Values

(attribute_name      VARCHAR2(30),
 attribute_type      VARCHAR2(106))

Usage Notes

This table function pipes out rows of type DM_MODEL_SIGNATURE. For information on Data Mining data types and piped output from table functions, see "Data Types".

Examples

Assume that you have built a classification model census_model using the Naive Bayes algorithm. You can retrieve the model details using Oracle SQL as follows:

SELECT attribute_name, attribute_type
  FROM TABLE(DBMS_DATA_MINING.GET_MODEL_SIGNATURE('census_model');

GET_MODEL_TRANSFORMATIONS Function

This function returns the transformation definitions associated with a model.

The data is transformed and the transformation definitions are embedded in the model during the model training (create) process. The transformations can be automatically generated, user-specified, or both.

Syntax

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

Parameters

Return Values

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

In addition to the transformation definitions, GET_MODEL_TRANSFORMATIONS returns the reverse transformation definitions. The reverse transformations are applied to the model attributes returned by the GET_MODEL_DETAILS functions. They are also applied to the target when a predictive model is applied. Reverse transformations enable model transparency, meaning that attribute information generated by the model is in its original, untransformed state, to the extent that it is possible.

GET_TRANSFORM_LIST Procedure

This function is used to convert a transformation specification from type DM_TRANSFORMS to type TRANSFORM_LIST.

DM_TRANSFORMS is used to output model transforms that are returned from the GET_MODEL_TRANSFORMATIONS function.

CREATE_MODEL and functions in the DBMS_DATA_MINING_TRANSFORM package accept TRANSFORM_LIST. To use them on the output of GET_MODEL_TANSFORMATIONS, you must first call GET_TRANSFORM_LIST to convert to TRANSFORM_LIST type.

Syntax

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

Parameters

IMPORT_MODEL Procedure

This procedure imports the specified data mining models from a dump file set that was created with EXPORT_MODEL or with the expdp export utility. Both IMPORT_MODEL and EXPORT_MODEL use Oracle Data Pump technology.

Syntax

DBMS_DATA_MINING.IMPORT_MODEL (
    filename             IN  VARCHAR2,
    directory            IN  VARCHAR2,
    model_filter         IN  VARCHAR2 DEFAULT NULL,
    operation            IN  VARCHAR2 DEFAULT NULL,
    remote_link          IN  VARCHAR2 DEFAULT NULL,
    jobname              IN  VARCHAR2 DEFAULT NULL,
    schema_remap         IN  VARCHAR2 DEFAULT NULL);

Parameters

'mymodel1'
'name IN (''mymodel2'',''mymodel3'')'

Examples

This example shows a model being exported and imported within the schema dmuser2. Then the same model is imported into the dmuser3 schema. The dmuser3 user has the IMPORT_FULL_DATABASE privilege.

SQL> connect dmuser2
Enter password: dmuser2_password
Connected.
SQL> select model_name from user_mining_models;
 
MODEL_NAME
------------------------------
NMF_SH_SAMPLE
SVMO_SH_CLAS_SAMPLE
SVMR_SH_REGR_SAMPLE

-- export the model called NMF_SH_SAMPLE to a dump file in same schema
SQL>EXECUTE DBMS_DATA_MINING.EXPORT_MODEL ('NMF_SH_SAMPLE_out', 'DATA_PUMP_DIR',
                            'name = ''NMF_SH_SAMPLE''');
-- import the model back into the same schema
SQL>EXECUTE DBMS_DATA_MINING.IMPORT_MODEL ('NMF_SH_SAMPLE_out01.dmp', 
                            'DATA_PUMP_DIR', 'name = ''NMF_SH_SAMPLE''');

-- connect as different user
-- import same model into that schema
SQL> connect dmuser3
Enter password: dmuser3_password
Connected.
SQL>EXECUTE DBMS_DATA_MINING.IMPORT_MODEL ('NMF_SH_SAMPLE_out01.dmp', 
                            'DATA_PUMP_DIR', 'name = ''NMF_SH_SAMPLE''',
                            'IMPORT', NULL, 'nmf_imp_job', 'dmuser2:dmuser3');

The following example shows user MARY importing all models from a dump file, model_exp_001.dmp, which was created by user SCOTT. The dump file is located in the file system directory mapped to a directory object called DM_DUMP. If user MARY does not have IMPORT_FULL_DATABASE privileges, IMPORT_MODEL will raise an error.

-- import all models
DECLARE
  file_name       VARCHAR2(40);
BEGIN
  file_name := 'model_exp_001.dmp';
  DBMS_DATA_MINING.IMPORT_MODEL(
                filename=>file_name,
               directory=>'DM_DUMP',                 schema_remap=>'SCOTT:MARY');
  DBMS_OUTPUT.PUT_LINE(
'DBMS_DATA_MINING.IMPORT_MODEL of all models from SCOTT done!');
END;
/

RANK_APPLY Procedure

This procedure ranks the results of an APPLY operation based on a top-N specification for predictive and descriptive model results. For classification models, you can provide a cost matrix as input, and obtain the ranked results with costs applied to the predictions.

Syntax

DBMS_DATA_MINING.RANK_APPLY (
      apply_result_table_name        IN VARCHAR2,
      case_id_column_name            IN VARCHAR2,
      score_column_name              IN VARCHAR2,
      score_criterion_column_name    IN VARCHAR2,
      ranked_apply_table_name        IN VARCHAR2,
      top_N                          IN INTEGER DEFAULT 1,
      cost_matrix_table_name         IN VARCHAR2 DEFAULT NULL,
      apply_result_schema_name       IN VARCHAR2 DEFAULT NULL,
      cost_matrix_schema_name        IN VARCHAR2 DEFAULT NULL);

Parameters

Usage Notes

You can use RANK_APPLY to generate ranked apply results, based on a top-N filter and also with application of cost for predictions, if the model was built with costs.

The behavior of RANK_APPLY is similar to that of APPLY with respect to other DDL-like operations such as CREATE_MODEL, DROP_MODEL, and RENAME_MODEL. The procedure does not depend on the model; the only input of relevance is the apply results generated in a fixed schema table from APPLY.

The main intended use of RANK_APPLY is for the generation of the final APPLY results against the scoring data in a production setting. You can apply the model against test data using APPLY, compute various test metrics against various cost matrix tables, and use the candidate cost matrix for RANK_APPLY.

The schema for the apply results from each of the supported algorithms is listed in subsequent sections. The case_id column will be the same case identifier column as that of the apply results.

Classification Models — NB, ABN, SVM

For numerical targets, the ranked results table will have the definition as shown:

(case_id       VARCHAR2/NUMBER,
prediction     NUMBER,
probability    NUMBER,
cost           NUMBER,
rank           INTEGER)

For categorical targets, the ranked results table will have the following definition:

(case_id       VARCHAR2/NUMBER,
prediction     VARCHAR2,
probability    NUMBER,
cost           NUMBER,
rank           INTEGER)

Clustering using k-Means or O-Cluster

Clustering is an unsupervised mining function, and hence there are no targets. The results of an APPLY operation contains simply the cluster identifier corresponding to a case, and the associated probability. Cost matrix is not considered here. The ranked results table will have the definition as shown, and contains the cluster ids ranked by top-N.

(case_id       VARCHAR2/NUMBER,
cluster_id     NUMBER,
probability    NUMBER,
rank           INTEGER)

Feature Extraction using NMF

Feature extraction is also an unsupervised mining function, and hence there are no targets. The results of an APPLY operation contains simply the feature identifier corresponding to a case, and the associated match quality. Cost matrix is not considered here. The ranked results table will have the definition as shown, and contains the feature ids ranked by top-N.

(case_id        VARCHAR2/NUMBER,
feature_id      NUMBER,
match_quality   NUMBER,
rank            INTEGER)

Examples

BEGIN
/* build a model with name census_model.
 * (See example under CREATE_MODEL)
 */ 

/* if training data was pre-processed in any manner,
 * perform the same pre-processing steps on apply
 * data also.
 * (See examples in the section on DBMS_DATA_MINING_TRANSFORM)
 */

/* apply the model to data to be scored */
DBMS_DATA_MINING.RANK_APPLY(
  apply_result_table_name       => 'census_apply_result',
  case_id_column_name           => 'person_id',
  score_column_name             => 'prediction',
  score_criterion_column_name   => 'probability
  ranked_apply_result_tab_name  => 'census_ranked_apply_result',
  top_N                         => 3,
  cost_matrix_table_name        => 'census_cost_matrix');
END;
/

-- View Ranked Apply Results
SELECT *
  FROM census_ranked_apply_result;

REMOVE_COST_MATRIX Procedure

Removes the default scoring matrix from a classification model.

Syntax

DBMS_DATA_MINING.REMOVE_COST_MATRIX (
      model_name   IN  VARCHAR2);

Parameters

Usage Notes

If the model is not in your schema, then REMOVE_COST_MATRIX requires the ALTER ANY MINING MODEL system privilege or the ALTER object privilege for the mining model.

Example

The Naive Bayes model NB_SH_CLAS_SAMPLE has an associated cost matrix that can be used for scoring the model.

SQL>SELECT *
      FROM TABLE(dbms_data_mining.get_model_cost_matrix('nb_sh_clas_sample'))
      ORDER BY predicted, actual;
 
ACTUAL     PREDICTED        COST
---------- ---------- ----------
0          0                   0
1          0                 .75
0          1                 .25
1          1                   0

You can remove the cost matrix with REMOVE_COST_MATRIX.

SQL>EXECUTE dbms_data_mining.remove_cost_matrix('nb_sh_clas_sample');

SQL>SELECT *
      FROM TABLE(dbms_data_mining.get_model_cost_matrix('nb_sh_clas_sample'))
      ORDER BY predicted, actual;

no rows selected

RENAME_MODEL Procedure

This procedure renames a mining model to a new name that you specify.

The model name is in the form [schema_name.]model_name. If you do not specify a schema, your own schema is used. For mining model naming restrictions, see "Mining Model Naming Restrictions".

Syntax

DBMS_DATA_MINING.RENAME_MODEL (
     model_name            IN VARCHAR2,
     new_model_name        IN VARCHAR2);

Parameters

Usage Notes

If an APPLY operation is using a model, and you attempt to rename the model during that time, the RENAME will succeed and APPLY will return indeterminate results.

Examples

Assume the existence of a model census_model. The following example shows how to rename this model.

BEGIN
  DBMS_DATA_MINING.RENAME_MODEL(
    model_name      => 'census_model',
    new_model_name  => 'census_new_model');
END;
/