Adaptive Bayes Network Algorithm
Adaptive Bayes Network (ABN) is an Oracle proprietary algorithm
that provides a fast, scalable, non-parametric means of extracting predictive information from data with respect to a target attribute.
This discussion describes the algorithm as it is implemented in the ODM Java interface; not all interface features may be exposed in Oracle Data Miner.
ABN supports three model types:
- SingleFeatureBuild
- MultiFeatureBuild
- NaiveBayesBuild
Each model type has different inputs and different results.
ABN predicts binary as well as multiclass targets.
Single Feature Build ABN Model Type
This is the only ABN model type that produces rules. This model type
builds a single multi-attribute feature.
The user specifies:
- The maximum depth (number of attributes) per multi-attribute feature
- The maximum number of attributes to consider during the build process (in rank order MDL correlation to target)
The algorithm determines
- Which attributes to put in the feature
- How many attributes to put in the feature up to the specified limit (Typically, the algorithm chooses only a very few attributes to include in a feature.)
Multi-Feature Build ABN Model Type
This ABN model type does not generate rules. This model builds a pruned Naive Bayes model and one or more multi-attribute features.
The user specifies:
- The maximum number of features to construct
- The maximum number of attributes to consider during the build process (in rank order MDL correlation to target)
- The maximum number of failures to accept a new feature before build termination (After each feature is constructed, a test determines whether the feature adds to or detracts from the accuracy of the overall model. If the new feature detracts from accuracy, the feature has failed acceptance.)
- The maximum depth (number of attributes) per multi-attribute feature
The algorithm determines the following:
- What attributes to put in what feature
- How many attributes to put in each feature up to specified limit
- Which features to include in the model
Naive Bayes Build ABN Model Type
This ABN model type does not generate rules. This model builds a pruned Naive Bayes model.
The user specifies:
- The maximum number of attributes to consider during the build process (in rank order MDL correlation to target)
The algorithm determines
- What attributes to include (based on their MDL rank)
This version of the ABN algorithm was introduced in an ODM 9.2 patch release; it is supported in ODM 10g. For detailed information about the algorithm, see the Javadoc documentation in the directory $ORACLE_HOME/dm/doc
on UNIX or %ORACLE_HOME%\dm\doc
on Windows.
Copyright © 2005, Oracle. All rights reserved.