EDOBE XDOM PMML Bedienerhandbuch PDF herunterladen (Seite 12)

Changes to the KMEANS Algorithm

 Clustering using Mahalanobis distance

 Normalized Euclidean distance

 Scoring with statistics of clusters and columns

 Automatic data normalization and standardization

 Enriched statistics

See “KMEANS algorithm” and “Enriched Statistics for Clustering Models” in the IBM SPSS

In-Database Analytics Developer's Guide for details of these new features.

In this release there is a behavior change to the KMEANS algorithm. By default, the 'nor-

malized Euclidean distance' is used by the KMEANS algorithm if the distance option is not

specified. In prior releases, 'Euclidean distance' is the default distance used.

Metadata Management For Analytic Models

The primary goal of the new Metadata Management feature is to provide an environment

for managing the analytic models created by the Netezza Analytics software. The imple-

mentation of the Metadata Management component is done on top of the existing database

system, using stored procedures and user-defined functions.

All analytics models created by the various Netezza Analytics functions (like DECTREE or

KMEANS) are registered in a catalog, and new administrative and other functions are of-

fered for model management. The Metadata Management system provides the following

features:

 List information about models

 Perform basic operations on models (for example, delete, copy, rename, update)

 Perform advanced operations on models ( for example print, PMML format, export)

 Security (grant and revoke privileges on models and model operations)

Note that this new feature is required by all algorithms that generate models. When you en-

able a database for Netezza Analytics using the script create_inza_db.sh, the database is

automatically prepared for the Metadata Management feature.

This new feature is described in more details in the “Metadata Management” section of the

IBM SPSS In-Database Analytics Developer's Guide.

Most models created using prior releases of Netezza Analytics can be registered in the

metadata catalog so that they can be used with Metadata Management. If model migration

is needed, it is done automatically. See the REGISTER_MODEL procedure in the IBM

SPSS In-Database Analytics Developer's Guide.

Limited PMML Support for Analytic Models

PMML (Predictive Model Markup Language) is defined by the Data Mining Group (DMG)

and is the widely accepted standard for the exchange of data mining models. Limited

PMML support is provided in this version. Support will be producer conformance for deci-

sion tree (classification), association rules, naïve Bayes, and k-means (clustering) models.

PMML support enables users to employ PMML-conformant model visualization tools, such

as the InfoSphere Warehouse visualizer, for model exploration. It also allows scoring of

Netezza Analytics models in DB2. The following algorithms have limited PMML support:

 Decision trees

1 2 ... 7 8 9 10 11 12 13 14 15 16 17 ... 27 28

Keine Kommentare

EDOBE XDOM PMML Bedienerhandbuch Seite 12