New software facilities

1
New software facilities

• Developed between
• Dec 2004 and Feb 2005

2
New implemented facilities

• Automatic and manual splits for
• Gini index of impurity.(used for classification)
• Variance measure of impurity.(used for
  regression)

3
Manual Split
4
Manual Split
5
Installer
6
General Options
7
CRT options
Obs For a large datasets use higher numbers
8
New implemented facilities

• Pruning using the Pessimistic Error Pruning
  algorithm (PEP) can be used only for pruning
  classification trees by Ross Quinlan (Quinlan,
  1987).
• Data transference from any node to another one.
• Disposal of sub nodes of any node, except leaves.

9
New implemented facilities

• Data filter and the selection of blocks of data
• Uses the set operations Union and
  Intersection.
• The variables can be either categorical or
  numerical.
• The comparison operators are , lt, lt, gt, gt and
  ltgt (different).
• The null keyword matches blank values.
• Silly-proof functions.
• The data block can be shown before transference.

10
Data filter and the selection of blocks of data
11
Spartacus ICU database, full grow tree (Target
outcome, 49 nodes)
12
Answer tree ICU database, full grown tree
(Target outcome, 11 nodes)
13
Spartacus ICU database, pruned tree (Target
outcome, 11 nodes)
14
Answer tree ICU database, pruned tree (Target
outcome, 1 node only )
15
Coming soon

• Tree misclassification tests.
• C4.5 implementation.
• More prune methods.
• Globalized numeric formats.
• Recalculation of the gain if a custom split is
  applied.
• Full support for missing values.

16
QA

• Any specific facility for the next meeting?
• See you by the end of May?
• http//evandro.org/calendar

17
(No Transcript)
18
PEP

• PEP (pessimistic error pruning) algorithm. PEP is
  a post-prune top down algorithm. It does not
  require a separate pruning set. PEP try to
  compensate  the overly optimistic estimates based
  on resubstitution error ( error based on the
  training sample tested with itself). These
  estimates are overly optimistic since error rates
  on training data are typically lower than on the
  test data. PEP attempts to get a more accurate
  error estimate by imposing correction for the
  binomial distribution by adding 0.5 to the number
  of errors associated with each node. Also,
  subtrees are not pruned only if their corrected
  error estimates is at least one standard error
  less than the estimated error of their root node.