A powerful non-parametric technique for classification
is CART, Classification and Regression Trees. (See
[White Book] Chapter 9.)
The S models library provides a convenient graphical
interface for some of the functions that operate on a
tree. However, this interface is command line driven
and also relies on internal routines which manipulate
the graphics device. With the recent changes in the
graphics device (briefly described in Multiple Graphics Devices)
and the use of the GUI library we can provide
a significantly more integrated, powerful and flexible
interface to the tree fitting routines in S.
We will describe a complete interface
to the tree routines that bring a user from
a dataset and variables to the comparison and investigation of
different trees and the properties of the fitted
trees. Also, we can contrast the fit from the tree with
that of a general linear model or other alternative techniques.
Each time a new data set is entered into the application
(either using the menus or via the invoking call),
a dialog box appears with two lists of the variables
in the dataset. This is used to create the formula
used by the tree() function.
The left list is an excusive list that allows the selection
of only one variable.
The second list allows the user to specify the predictors
in the model.
If a variable is selected
in one list is not available in the other list.
These can be either scrollable lists or check and radio
boxes.
These lists may be layed out horizontally or side by side.
They may also be stacked on top of a data browser (spreadsheet
like table) so the user can see which variable names
relate to which values.
or
Note that the user can't select a variable in one that is
already selected in the other list.
The code must prohibit this. It should also color
those buttons available in each list appropriately
or make them insensitive.
When the user hits the Fit button, the main work window appears.
This can be created in a separate thread while the user is
selecting the model inputs.
The components of this window are the
- menu bar,
- Data
- New Data Set
- Clone Window
- Print Tree
- Predict using some data and the given tree.
- Apply user specified function to a tree.
- Show Current Data
- Confusion matrix.
- Deviance versus Size plot
- Cross Validation
- Views
- List of Cards in this window
- List of Windows in this application
- Options
- Node plots in Panel or Window
- Select
- Ancestors
- Descendants
- Select All
- Unselect All
- Delete
- As New Tree
- Drag to graft to another part of the tree
- Help
- Local to this application
- Index/Table of Contents
- model description,
Click on this to edit the model formula - dependent
and independent variables.
- Graphics display with the current fitted tree
which displays the tree.
Each node has a popup menu and an action associated
with pressing the button
- browser - window containg text about this node
- burl - competing splits at this point
- identify - if data window is open, the
corresponding observations are highlighted
- hist - gives a selection of variables to observe
for the left and right sides of the subtree
with this node as its root.
Output is displayed in a new window
or in a new panel below this tree.
Controlled by boolean specifiable in options.
In the data browser for this tree,
allow for path.tree() to highlight the splits
that get an observation to the node in the tree.
Last modified: Sun Feb 16 12:45:36 1997