by

What's New in CART® 4.0 Pro

Salford Systems' latest release of our flagship data-mining product, CART 4.0 PRO, contains many new features. CART 4.0 PRO is available for Windows 95 and 98, Windows NT 4.0, Windows 2000, and all major versions of UNIX and Linux.

Automatic Generation of Customized Reports and Publishing to the Web

Customized reports now can be generated automatically for any tree no matter when it was grown. Options control which components of individual CART analyses (tree diagram, gains charts, importance rankings, etc.) are included in each report. If several trees have been grown, a summary report table can be produced that describes each tree in one line of the table. The report can be saved as a Rich Text Format (.rtf) document and then published to Web (as .html).

Class Labels

You now can assign 32-character labels to levels of categorical variables. Labels can be added interactively to any variable including the target. Tree diagrams, CART text output and displays can use these labels when referring to levels of either target or predictor variables. Once entered the labels can be exported to any other analysis or tree; labels for categorical variables with many levels can be applied via command script and need not be entered interactively.

Case Weights

Case weights now can be used with any CART splitting rule or tree-combining method. (CART formerly treated each observation equally, i.e., with an effective case weight of 1.0.) Case weights, which may take on fractional values as well as whole numbers, are stored in a variable in the dataset and typically vary from observation to observation.

Penalties for Predictor Variables

CART now offers three ways to "penalize" (scale down) the improvement computed for a predictor variable, so that it is less likely to be chosen as the primary splitter. Penalties can be used to reflect the cost of a predictor's acquisition or the preference to not use a certain predictor unless it is unusually effective. The user can specify penalties on a specific variable from none through light to heavy. In addition, exact numerical penalty values can be specified or penalties can be imposed on all predictor variables in proportion to the degree that they are missing.

Penalties for High-Level Categorical Variables

A new method is provided to use effectively a predictor variable with many levels (such as zip code, which may have 3000 distinct levels) without giving such a predictor an undue advantage. Spurious splits are greatly reduced by ensuring that a high-level categorical predictor has no inherent advantage over a continuous variable that has a unique value for each record.

Combined Trees

CART's handling of committees of trees (bagging, ARCing) has been improved to make this technology more accessible to the average CART user. In addition, the user now can learn details about individual trees used in the committee. In some cases these details can allow users to draw robust conclusions that are not otherwise apparent.

© Copyright 2002 Salford-Systems Inc.