Oracle Data Mining implements a variety of data mining algorithms inside the Oracle relational database. These implementations are integrated right into the Oracle database kernel, and operate natively on data stored in the relational database tables. This eliminates the need for extraction or transfer of data into standalone mining/analytic servers. The relational database platform is leveraged to securely manage models and efficiently execute SQL queries on large volumes of data. The system is organized around a few generic operations providing a general unified interface for data mining functions. These operations include functions to create, apply, test, and manipulate data mining models. Models are created and stored as database objects, and their management is done within the database - similar to tables, views, indexes and other database objects.
In data mining, the process of using a model to derive predictions or descriptions of behavior that is yet to occur is called "scoring". In traditional analytic workbenches, a model built in the analytic engine has to be deployed in a mission-critical system to score new data, or the data is moved from relational tables into the analytical workbench - most workbenches offer proprietary scoring interfaces. ODM simplifies model deployment by offering Oracle SQL functions to score data stored right in the database. This way, the user/application developer can leverage the full power of Oracle SQL - in terms of the ability to pipeline and manipulate the results over several levels, and in terms of parallelizing and partitioning data access for performance.
Models can be created and managed by one of several means. (Oracle Data Miner) is a graphical user interface that steps the user through the process of creating, testing, and applying models - approximating the CRISP-DM methodology. Application and tools developers can embed predictive and descriptive mining capabilities using PL/SQL or Java APIs. Business analysts can quickly experiment with, or demonstrate the power of, predictive analytics using Oracle Spreadsheet Add-In for Predictive Analytics, a dedicated Microsoft Excel adaptor interface. ODM offers a choice of well known machine learning approaches such as Decision Trees, Naive Bayes, Support vector machines, Generalized linear model (GLM) for predictive mining, Association rules, K-means and Orthogonal Partitioning Clustering (see O-Cluster paper below), and Non-negative matrix factorization for descriptive mining. A minimum description length based technique to grade the relative importance of an input mining attributes for a given problem is also provided. Most Oracle Data Mining functions also allow text mining by accepting Text (unstructured data) attributes as input.
Oracle Data Mining was first introduced in 2002 and its releases are named according to the corresponding Oracle database release:
Oracle Data Mining is a logical successor of the Darwin data mining toolset developed by Thinking Machines Corporation in the mid-1990s and later distributed by Oracle after its acquisition of Thinking Machines in 1999. However, the product itself is a complete redesign and rewrite from ground-up - while Darwin was a classic GUI-based analytical workbench, ODM offers a data mining development/deployment platform integrated into the Oracle database, along with the GUI.
As of release 11gR1 Oracle Data Mining contains the following data mining functions:
Most Oracle Data Mining functions accept as input one relational table or view. Flat data can be combined with transactional data through the use of nested columns, enabling mining of data involving one-to-many relationships (e.g. a star schema).
Oracle Data Mining distinguishes numerical, categorical, and unstructured (text) attributes. The product also provides utilities for data preparation steps prior to model building such as outlier treatment, discretization, normalization and binning (sorting in general speak)
Oracle Data Mining can be accessed using Oracle Data Miner a GUI “client” that provides access to the data mining functions and structured templates called Mining Activities that automatically prescribe the order of operations, perform required data transformations, and set model parameters. The user interface also allows the automated generation of Java and/or SQL code associated with the data mining activities. The Java Code Generator is an extension to Oracle JDeveloper. There is also an independent interface: the Spreadsheet Add-In for Predictive Analytics which enables access to the Oracle Data Mining Predictive Analytics PL/SQL package from Microsoft Excel.
Oracle Data Mining provides a native PL/SQL package (DBMS_DATA_MINING) to create, destroy, describe, apply, test, export and import models. The code below illustrates a typical call to build a classification model:
BEGINDBMS_DATA_MINING.CREATE_MODEL (model_name => 'credit_risk_model',function => DBMS_DATA_MINING.classification,data_table_name => ‘credit_card_data',case_id_column_name => 'customer_id',target_column_name => ‘credit_risk',settings_table_name => 'credit_risk_model_settings');END;
where 'credit_risk_model' is the model name, built for the express purpose of classifying future customers' 'credit_risk', based on training data provided in the table 'credit_card_data', each case distinguished by a unique 'customer_id', with the rest of the model parameters specified through the table 'credit_risk_model_settings'.
Oracle Data Mining also supports a Java API consistent with the Java Data Mining (JDM) standard for data mining (JSR-73) for enabling integration with web and J2EE applications and to facilitate portability across platforms.
As of release 10gR2, Oracle Data Mining contains built-in SQL functions for scoring data mining models. These single-row functions support classification, regression, anomaly detection, clustering, and feature extraction. The code below illustrates a typical usage of a classification model:
SELECT customer_nameFROM credit_card_dataWHERE PREDICTION (credit_risk_model USING *) = 'LOW' AND customer_value = 'HIGH';
The PL/SQL package DBMS_PREDICTIVE_ANALYTICS automates the data mining process including data preprocessing, model building and evaluation, and scoring of new data. The PREDICT operation is used for predicting target values classification or regression while EXPLAIN ranks attributes in order of influence in explaining a target column feature selection. The new 11g feature PROFILE finds customer segments and their profiles, given a target attribute. These operations can be used as part of an operational pipeline providing actionable results or displayed for interpretation by end users.