12.2.2 Major Classes and Associations

   The CWM Data Mining metamodel consists of seven conceptual areas: A core Mining metamodel (upon which the other areas depend), and metamodels representing the data mining subdomains of Clustering, Association Rules, Supervised, Classification, Approximation, and Attribute Importance. Each area is represented by the metamodel packages shown in the diagram below.

   DataMining (from Analysis)

   <<metamodel>> <<metamodel>> <<metamodel>> AssociationRules

   Figure 12-1 CWM Data Mining Metamodel

   Collectively, the collection of Data Mining packages provide the necessary abstractions to model generic representations of data mining models (i.e., mathematical models produced or generated by the execution of data mining algorithms).

   Included are representations of data mining tasks and models, as well as other entities (such as category matrix) that are common across most data mining applications and tools, as well as their relationships to each other and their mappings to technical metadata.

   The Mining Core package consists of common Data Mining abstractions that are fundamental to, and reused by, the major conceptual areas. In particular, this package contains several basic packages that are required to implement the CWM Data Mining interfaces. It is required that at least this package and one more Data Mining package be implemented for compliance. The packages forming the Mining Core are shown in the next diagram.

   <<metamodel>> MiningCore (from DataMining)

   Figure 12-2 CWM Data Mining Metamodel: Mining Core Package

   The following subsections describe the content of each component package of the MiningCore. This is subsequently followed by subsections describing each of the major conceptual area packages.

   12.2.2.1 Mining Function Settings

   settings logicalData

   Logical Data

   Figure 12-3 CWM Data Mining Metamodel: Mining Function Settings

   This package defines the objects that contain parameters specific to mining functions. The separation of mining functions from mining algorithms enables the user to specify the type of the desired result without being concerned with a particular algorithm. The Mining Function Settings metamodel is illustrated above.

   Mining FunctionSettings (MFS) is the superclass of all other function settings classes. An MFS instance references a set of MiningAttributes, aggregated by a LogicalData instance. The AttributeUsage set defines how each of the MiningAttributes will be used by the Mining Algorithm.

   12.2.2.2 Mining Model

   +modelLocation

   MiningModel

   SignatureAttribute

   (from Core) (fromMiningFunctionSettings)

   Figure 12-4 CWM Data Mining Metamodel: Mining Model

   This package defines the basic Mining Model from which all model objects inherit as the result of a mining build task. The Mining Model metamodel is illustrated above.

   Each MiningModel has a signature that defines the characteristics of the data required by the model.

   12.2.2.3 Mining Result

   ModelElement

   MiningResult

   Figure 12-5 CWM Data Mining Metamodel: Mining Result

   This package defines the basic MiningResult object from which all result objects inherit as the result of a specific mining task (other than build).

   12.2.2.4 Mining Data

   This package defines the objects that describe the input data, the way the input data is to be treated, and the mapping between the input data and internal representation for which mining algorithms can understand.

   PhysicalData effectively references and instance of a class or subclass (e.g., Table, file, etc.). This allows JDM to leverage the various row/column format data representation expressible in CWM.

   Mining Data metaclasses representing the concepts of physical data are illustrated in Figure 12-6. Logical data metaclasses are illustrated in Figure 12-7. Attribute assignment and attribute usage metaclasses are illustrated in two subsequent diagrams ( Figure 12-8 and Figure 12-9, respectively).

   Finally, metaclasses used to model the matrix representation and taxonomy of mining data are presented in Figure 12-10, Category Matrix, and Figure 12-11, Category Taxonomy, respectively.

   ModelElement (from Core)

   Figure 12-6 CWM Data Mining Metamodel: Physical Data

    Figure 12-6 illustrates those elements of the Mining Data metamodel used to model physical data, whereas the following diagram shows those elements facilitating the logical modeling of data.

   Class (from Core)

   Attribute (from Core)

   LogicalData

   LogicalAttribute

   numericalProperties

   categoricalProperties

   {ordered}

   taxonomy

   Figure 12-7 CWM Data Mining Metamodel: Logical Data

    Figure 12-7 contains objects that represent how physical data should be interpreted, logically by the mining algorithm.

   A LogicalAttribute can be categorical, numerical, or both, depending on its usage. Categorical attributes that have ordered category values are created as ordinal attributes.

   12-8 Common Warehouse Metamodel, v1.1 March 2003

   AttributeAssignmentSet

   set

   MiningAttribute

   attrAssignmnet orderIdAttribute Attribute{ordered} (from Core)

   Pi votAttributeAssi gnment DirectAttributeAssignment

   pivotAttrAssignment

   Attribute (from Core)

   nameAttribute

   SetAttri buteAssignment

   setAttrAssignment setAttrAssignment

   setIdAttribute

   valueAttribute ReversePivotAttributeAssignment

   setIdAttribute

   reversePivotAttrAssignment

   0.

   Attribute selectorAttribute

   (from Core)

   Figure 12-8 CWM Data Mining Metamodel: Attribute Assignment

   • Figure 12-8 illustrates metaclasses that enable mapping physical data attributes to logical data mining attributes. The following attribute assignments are supported:

• Direct assignment: A direct mapping between a mining attribute and a physical attribute.

• Pivot assignment: A mapping where the input data is in transactional format; each of the logical attributes occurring in a pivoted table is mapped to the three physical columns, presumably the same ones every time.

• Reverse pivot assignment: A mapping where the input data is in 2D format; the transformed input data contains set valued attributes; the sets are represented by enumerating the set elements based on the selection function.

• Set assignment: A mapping between a set valued mining attribute and a set of attributes in the physical data.

   March 2003 OMG-CWM, v1.1: Organization of the Data Mining Metamodel

   Class(from Core)

   AttributeUsageSet

   Feature (from Core) /owner

   /feature

   attribute

   usage

   Figure 12-9 CWM Data Mining Metamodel: Attribute Usage

    Figure 12-9 illustrates metaclasses that enable specification of how a mining attribute should be used, interpreted, or preprocessed (e.g., mining value or outlier/invalid value treatment).

   CategoryMatrix

   categoryMatrix category

   CategoryMatrixObject

   categoryMatrix

   matrixTaable

   CategoryMatrixTable source Class

   matrixTable

   entry

   col umnIndex columnAttribute

   (from Core) valueAttribute

   Figure 12-10 CWM Data Mining Metamodel: Category Matrix

    Figure 12-10 illustrates the metaclasses that generalize a complex object used to represent a cost matrix (a model build input) or a confusion matrix (a model test result). Two representations are supported:

• Java objects (CategoryMatrixObject)

• Table based (CategoryMatrixTable)

   March 2003 OMG-CWM, v1.1: Organization of the Data Mining Metamodel

   CategoryTaxonomy taxonomy CategoryMapcategoryMap

   taxonomy

   CategoryMapObject

   mapTable mapTable

   mapObject

   mapTable

   entry

   CategoryMapObjectEntry

   entry

   parent

   Category

   (from Core) graphIdAttributerootCategory

   Figure 12-11 CWM Data Mining Metamodel: Category Taxonomy

    Figure 12-11 also illustrates the metaclasses that enable representing a taxonomy as a directed acyclic graph (DAG). Two representations are supported::

• Java Object (CategoryMapObject)

• table-bound (CategoryMapTable).

   Mining Task

   This package defines the objects that are related to mining tasks. A MiningTask object represents a specific mining operation to be performed on a given data set (i.e., physical data). Figure 12-12 illustrates the basic Mining Task metamodel.

   Transformation (from Transformation)

   MiningTransformation ModelElement(fromCore)

   transformation

   procedure

   MiningTask

   miningTaskMini ngModel inputModel miningTask

   (fromMiningModel)

   inputData

   PhysicalDatamodelAssignment

   (from MiningData) AttributeAssignmentSet (from MiningData)

   Figure 12-12 CWM Data Mining Metamodel: Mining Task

    Figure 12-12 illustrates Mining Task as referenced by a Mining Transformation. A Mining Task maps physical data to a model signature (when applicable; for example, lift, test, etc.) using the Attribute Assignment set.

   Min ingTask

   MiningBuildTask buildTask

   validationData

   buildTask

   buildTask buildTaskbuildTask

   (from MiningData)

   validationAssignmentresultModel miningSettings

   settingsAssignment

   Min ingModel MiningFunctionSettings

   AttributeAssignmentSet

   (from MiningModel) (from MiningFunctionSettings)

   (from MiningData)

   Figure 12-13 CWM Data Mining Metamodel: Mining Build Task

   Model elements comprising the Mining Build Task are shown in Figure 12-13 . The modeling of the application of output and the computation of the result of an application of a data mining model to (new) data are illustrated in Figure 12-14 and Figure 12-15 , respectively.

   MiningApplyOutput

   MiningAttribute {ordered}

   item

   ApplySourceItem

   ApplyContentItem

   ApplyProbabilityItem ApplyScoreItem ApplyRuleIdItem

   Figure 12-14 CWM Data Mining Metamodel: Apply Output

    Figure 12-14 illustrates metaclasses that enable defining the content of an Apply task. This includes source items; for example, keys, or specific content of apply (data scoring using a model).

   An apply output may contain multiple source and content items.

   MiningTask

   MiningApplyOutput AttributeAssignmentSet (from MiningData)

   Figure 12-15 CWM Data Mining Metamodel: Mining Apply Task

    Figure 12-15 illustrates metaclasses that allow specification of an apply task. The apply task requires a model, physical data, apply output, and an attribute assignment set.

   Entry Point

   This package defines the top-level objects of DataMining package which can be used as entry point in application programming. This is illustrated in Figure 12-16 .

   Package (from Core)

   CatalogLogicalData(from MiningData)

   (from MiningResult)

   logicalData

   Schema

   schema

   schemaschema

   auxOobjects

   schema

   0..*

   MiningModel auxiliaryObject

   task

   miningFunctionSettings

   MiningTask

   taxonomy

   CategoryTaxonomy (from MiningFunctionSettings) (from MiningTask) attributeAssignmentSet (from MiningData) AttributeAssignmentSet (from MiningData)

   Figure 12-16 CWM Data Mining Metamodel: Entry Point

   Clustering

   This package contains the metamodel that represents clustering functions, models, and settings. The Clustering metamodel is illustrated in Figure 12-17 . It contains attribute usage and function settings, subclasses that are specific to the Clustering function.

   March 2003 OMG-CWM, v1.1: Organization of the Data Mining Metamodel

   AttributeUsage (from MiningData)

   ClusteringAttributeUsage attributeComparisonFunction : AttributeComparisonFunction similarityScale : Double / comparisonMatrix : CategoryMatrix

   attributeUsage

   comparisonMatrix

   CategoryMatrix (from MiningData)

   MiningFunctionSettings(from MiningFunctionSettings)

   ClusteringFunctionSettings maxNumberOfClusters :Integer minClusterSize : Integer = 1 aggregationFunction : AggregationFunction

   Figure 12-17 CWM Data Mining Metamodel: Clustering

   Association Rules

   This package contains the metamodel that represents the constructs for frequent itemset, association rules and sequence algorithms. The Association Rules metamodel is illustrated in Figure 12-18 .

   MiningFunctionSettings (fromMiningFunctionSettings)

   FrequentItemSetFunctionSettings

   settings exclusion Category

   AssociationRulesFunctionSettings SequenceFunctionSettings

   Figure 12-18 CWM Data Mining Metamodel: Association Rules

   12.2.2.5 Supervised

   This package contains the metamodel that represents the constructs for supervised learning algorithms. The Approximation, Attribute Importance, and Classification packages must implement this package. Figure 12-19 illustrates the Supervised metamodel. It contains test and lift tasks, test and lift results, and a common superclass for supervised function settings.

   MiningTask (from MiningTask)

   MiningResult(from MiningResult)

   MiningTestTask MiningTestResult

   testTask

   liftAnalysis

   positiveTargetCategory

   MiningFunctionSettings(from MiningFunctionSettings)

   liftAnalysis

   point

   LiftAnalysisPoint

   SupervisedFunctionSettings

   Figure 12-19 CWM Data Mining Metamodel: Supervised

   Classification

   This package contains the metamodel that represents classification function, models, and settings.

   .

   SupervisedFunctionSettings (from Supervised)

   ClassificationFunctionSettings

   (from MiningData)

   Figure 12-20 CWM Data Mining Metamodel: Classification Function Settings

    Figure 12-20 represents the model for Function Settings, while Figure 12-21 illustrates those model elements used to represent Attribute Usage that can include prior probability specification. Figure 12-22 shows that portion of the Classification metamodel modeling Classification Test tasks, results, and apply output.

   AttributeUsage (from MiningData)

   ClassificationAttributeUsage

   usage

   priors

   positiveCategory 1..*1..*

   Category(from MiningData)

   priors

   prior

   Figure 12-21 CWM Data Mining Metamodel: Classification Attribute Usage

   MiningTestTask (from Supervised)

   MiningTestResult (from Supervised)

   ClassificationTestTask testTask testResult

   testResult

   confusionMatrix

   ApplyOutputItem (from MiningTask)

   CategoryMatrix (from MiningData)

   ApplyTargetValueItem

   (from MiningData)

   Figure 12-22 CWM Data Mining Metamodel: Classification Test and Result

   Approximation

   This package contains the metamodel that represents the constructs for approximation modeling (also known as regression). The metamodel is shown in Figure 12-23 .

   March 2003 OMG-CWM, v1.1: Organization of the Data Mining Metamodel

   MiningTestTask

   MiningTestResult (from Supervised)

   SupervisedFunctionSettings (from Supervised)

   ApproximationFunctionSettings

   Figure 12-23 CWM Data Mining Metamodel: Approximation

   Attribute Importance

   This package contains the metamodel that represents the constructs for attribute importance (also known as feature selection) model. This metamodel is illustrated in Figure 12-24 .

   SupervisedFunctionSettings (from Supervised)

   AttributeImportanceSettings

   Figure 12-24 CWM Data Mining Metamodel: Attribute Importance