9
Oracle Data Mining Scoring Engine

Some data-mining enabled applications have models that are developed on one system and then deployed to other (production) systems. Production applications often need only to apply models built elsewhere. The Oracle Data Mining Scoring Engine supports scoring data (applying models) using models created elsewhere.

The Scoring Engine allows customers to limit the Oracle Data Mining (ODM) functionality available within their scoring applications to ensure that compute-intensive operations such as model building are not performed on systems where Scoring Engine is installed.

9.1 Oracle Data Mining Scoring Engine Features

The ODM Scoring Engine supports operations for preparing data as required from the build process, importing a model, and applying a model to data. All transformation functionality is included in the Scoring Engine. All functionality provided in the Scoring Engine behaves exactly as the full system.

You cannot build models using the Scoring Engine.

9.2 Data Mining Scoring Engine Installation

"Data Mining Scoring Engine" is a custom install option for Oracle Data Mining; select this option to install the ODM Scoring Engine.

9.3 Scoring in Data Mining Applications

A single model can be used to score large volumes of data, often in multiple geographically distributed application settings. Data analysis and model building might be performed by a small group of data mining experts using data from a centralized data warehouse. However, the model, can be used by a much larger number of applications working with data at geographically dispersed sites using local data. Local data may consist of millions of records representing customers; therefore, it can make sense to move the model to where the data is.

In real-time applications such as call centers, models are often built in one environment and used in another. There may be one machine dedicated to model building, using large volumes of data to produce models on a daily basis. Several other machines may be dedicated to real-time scoring, receiving new models to support, for example, the call center application. Call center representatives collect information from callers; the collected information is then used to obtain predictions or recommendations for that particular caller in real time. Scoring in real time often requires that the model is moved to where the data is.

9.4 Moving Data Mining Models

Oracle Data Mining supports two ways to move models from one schema or debase instance to another:

• PMML export and import in the Java interface
• Native export and import in the PL/SQL interface

9.4.1 PMML Export and Import

You can transport Naive Bayes and Association models between ODM instances using the Export task followed by the Import to import. The PMML string produced by the Export task can be moved via file transport mechanisms for import at a target ODM instance or Scoring Engine instance. The ODM Export Task places the PMML string in a table in a cell of type CLOB, The Import Task reads the PMML string from a table cell of type CLOB. The sample program PMMLDemo.java illustrates exporting and importing a Naive Bayes model.

9.4.2 Native ODM Export and Import

In a narrow sense, native export and import implies moving data mining model out of and into Oracle databases using Oracle proprietary facilities.

Native export and import are currently supported for data mining models built using the ODM PL/SQL interface only, namely those models created using the DBMS_DATA_MINING package.

Naive model export and import is based on Oracle Data Pump technology.

Native export is supported at three different levels, as follows:

• When a full database is exported using the Oracle Data Pump Export Utility (expdp), all data mining models in the database are exported.
• When a schema is exported using the Oracle Data Pump Export Utility (expdp), all data mining models in the schema are exported.
• An ODM user can export one or more specific models in a schema using the DBMS_DATA_MINING.export_model procedure.

Native import is also supported in all scenarios. Using the dump file set produced by the Oracle Data Pump Export Utility (expdp), an ODM user can run the Oracle Data Pump Import Utility (impdp) to import all data mining models contained in the dump file set.

ODM users can import a specific model from the dump file set using DBMS_DATA_MINING.import_model procedure.

For more information about native export and import, see the Oracle 10g Data Mining Administrator's Guide and the DBMS_DATA_MINING chapter in the PL/SQL Packages and Types Reference.

9.5 Using the Oracle Data Mining Scoring Engine

Suppose that the application builds a model on one system named BLDSYS and applies the model on a different system named SCORESYS.

BLDYS must have ODM installed. ODM supports all data mining activities (building models, testing models, applying models, etc.). SCORESYS can have the ODM Scoring Engine installed. It will not be possible to build models on SCORESYS, but it will be possible to apply the model. Note that SCORESYS could have a full ODM product installation.

The following processing takes place:

1. Build models on BLDYSYS. Select the appropriate model to deploy.
2. Export the model using the ODM PL/SQL interface (or use the Java interface to export Naive Bayes and Association models).
3. Copy the dump file generated by model export to SCORESYS. If you used PMML export, copy the PMML XML string to SCORESYS and export the string from the table cell.
4. Import the model on SCORESYS using the corresponding interface that you used to export it.
5. Apply the model to data on SCORESYS.