Neural Networks Products
DBMS, Data Mining Solutions Supplement
Despite their complexity ý or possibly because of it -- neural networks are very powerful data mining tools. They can be found today as components of suite products, i.e., products that have several data mining tools; as well as stand-alone products. Unless otherwise noted, each neural network is suitable for both classification and regression.
Cognos Inc: 4Thought (Windows 3.1/95). The product is not designed to do classification, although a user can apply it to classification by making appropriate conversions. It has good feedback during training and a large set of interpretive and sensitivity tools to help increase user understanding of and confidence in a model. It includes features for modeling time series data.
Group 1 Software: Model 1 (Windows 95/NT). Model 1 includes the neural net algorithm from Unica Technologyýs PRW (see below). It has been integrated into an interface designed specifically for predicting customer response to direct mail, advertisements, and other marketing campaigns.
HNC Software Inc.: DataBase Mining Marksman (Windows NT). A hardware and software solution, DataBase Mining Marksman includes a special board containing 16 processors that operate in parallel. It has an excellent pre-processing facility for data cleanup. Post-processing analysis oriented to customer response applications can compute lift and ROI. Requires Microsoft Excel.
Integral Solutions Ltd.: Clementine (UNIX, VMS, Windows NT). The Clementine tool set has several algorithms, including a neural net. Clementine has a unique visual programming interface that, by modeling the data flow in a data mining exercise, permits multiple algorithms to be linked to each other, to data preparation steps and to a variety of outputs. The neural net algorithms provide both simple and expert modeling levels with access to numerous parameters.
IBM: The Intelligent Miner (AIX, MVS, OS400). A product suite that includes a neural net for classification only. This product is easier to use than most neural nets because algorithm controls are hidden from the user, but expert users will be frustrated by the lack of control. The algorithm is scalable on parallel processing architectures.
NCR Corp.: Knowledge Discovery Workbench (KDW) is an implementation of Integral Solutions Ltd.ýs Clementine (see above).
NeoVista Software Inc.: DecisionNet (UNIX on DEC, HP and Sun). Decision Net is the neural network component of Decision Series, and is an implementation of a back propagation algorithm. The algorithm is scalable on parallel processing architectures.
NeuralWare Inc.: NeuralWorks Predict (Windows 3.1/95/NT, UNIX on DEC, HP, IBM RS6000, Silicon Graphics and Sun). NeuralWorks Predict uses adaptive gradient or Kalman learning techniques rather than the more common back propagation. The adaptive gradient technique supports dynamic searching not just through weights but also through neural net topology. Predict supports four different levels of expertise from novice through expert, with additional controls becoming available at each level. Requires Microsoft Excel in the Windows environment.
SAS Institute Inc.: SAS Data Mining Solution (Windows 3.1/95/NT, OS2, Macintosh, any UNIX, IBM MVS/VM/VSE, DEC VMS). SAS Data Mining Solution is an optional add-on to the SAS system, and includes the SAS Neural Network Application. The algorithm has numerous user settable parameters including topology, activation functions and error functions.
SPSS Inc.: SPSS Neural Connection (Windows 3.1/95/NT). SPSS Neural Connection is an optional SPSS tool, and offers neural networks for model generation. Data set size limitations such as a maximum of 10000 training cases, may restrict its application.
Thinking Machines Corp.: Darwin (Windows 95/NT client, HP/UX, AIX UNIX and Sun Solaris servers). The Darwin suite of products includes StarNet for generating predictive models with neural networks. The The algorithm is scalable on parallel processing architectures.
Unica Technologies Inc.: Pattern Recognition Workbench (Windows 3.1/95/NT). Pattern Recognition Workbench (PRW) has numerous user controllable parameters. Users can select from a variety of training algorithms, and can also select from various searching algorithms.