Modifications of C4.5 since book was published:
--------------------------------------------------------------------------------
(1) 17 August 1992: fixed bug in prunerule.c
In routine Satisfies about line 434:
moved statement
t->Outcome = -1;
to before the for loop
--------------------------------------------------------------------------------
(2) 2nd Feb 1993: fixed errors reported by Dick Jackson
c4.5rules.c line 34: changed ';' to ','
getnames.c: moved CopyString() declaration to head
--------------------------------------------------------------------------------
(3) 19th June 1993: fixed error reported by Guillermo Irisarri
ANSI C doesn't like "exit()" with no args in average.c, xval-prep.c
--------------------------------------------------------------------------------
(4) 5th July 1993: fixed bug in c4.5rules reported by Ray Mooney
SaveRules() was invoked before EvaluateRulesets(), but the latter
can delete globally unhelpful rules. SaveRules() was moved to
after evaluation of rules on training data
(Note: this change affects only the use of consultr with the
saved rules; experimental results are unaltered.)
--------------------------------------------------------------------------------
(5) 13th July 1993: changed rules.c to improve printing with -s option
When tests on discrete attributes use value groups, the standard
form of test is
"<attribute> in {<value>, <value>, ...}".
If there is only one value, this should appear as
"<attribute> = <value>".
This has already been changed in trees; a similar change has now
been made to function PrintCondition() in rules.c
--------------------------------------------------------------------------------
(6) 28th July 1993: killed very large confusion matrices
confmat.c line 19: added copout if number of classes > 20
--------------------------------------------------------------------------------
(7) 9th September 1993: fixed problems notified by Mike Jankulak.
* Added checks for reasonable parameter values in c4.5, c4.5rules.
Check in GetNames() for discrete N: N must be at least 2.
* consult, consultr don't work with attributes of type discrete N !
Added routines in trees.c to save and restore values of attributes
of this type when saving / reading trees.
Modified rules.c to invoke these routines when saving / reading
rulesets.
NOTE: old .tree, .unpruned and .rules files must be regenerated
if they are to be used by the modified programs.
--------------------------------------------------------------------------------
(8) 3rd November 1993; problem notified by Jason Catlett
c4.5rules prints an incorrect confusion matrix for the training
set when rules are dropped. Altered testrules.c.
--------------------------------------------------------------------------------
(9) 21st December 1993; tidying up only
Changed definition of Log() in defns.i so that argument of log()
is guaranteed float.
--------------------------------------------------------------------------------
(10) 5th February 1994; problem notified by George John
Calculation of Gain in build.c can be negative rather than zero
due to FP rounding. Changed tests "Gain[Att] >= 0" to
"Gain[Att] > -Epsilon".
--------------------------------------------------------------------------------
(11) 25th May 1994; problem notified by Ronny Kohavi
Similar problem in info.c with -g option. Changed test
"ThisGain > 0" to "ThisGain > -Epsilon".
--------------------------------------------------------------------------------
(12) 30th May 1994; tidying up
Removed explicit Outcomes field from rules. This simplifies
the code somewhat with little decrease in efficiency.
--------------------------------------------------------------------------------
(13) 18th July 1994; problem notified by Ronny Kohavi
Average gain evaluated incorrectly when all attributes have
many discrete values. In build.c, introduced MultiVal to check
for this contingency.
--------------------------------------------------------------------------------
(14) 18th-20th July 1994; modifications to siftrules.c
(a) Changed coding of exceptions:
* added cost of encoding total number of errors to cost of
identifying false positives and false negatives.
* applied penalty to non-representative theories as described
in my ML'94 paper.
(b) Introduced a new form of local greedy search for finding
good subsets when there are more than 10 rules. This is
faster than simulated annealing and replaces it as the default:
simulated annealing is still available via a new option -a.
--------------------------------------------------------------------------------
*********************** Release 6 July 1994 *********************************
--------------------------------------------------------------------------------
(15) 11th August 1994; bug reported by KaiMing Ting and Zijian Zheng
In subset.c, DiscrKnownBaseInfo() can be called when KnownItems = 0.
Trapped such calls.
--------------------------------------------------------------------------------
(16) 21st January 1995; bug reported by Tom Fawcett
Very large trees can cause the short int in TreeSize to overflow.
Changed to int.
--------------------------------------------------------------------------------
(17) 6th April 1995; bug reported by Ronny Kohavi
Exit status not being set properly. Modified the following:
c4.5.c, c4.5rules.c, consult.c, consultr.c.
--------------------------------------------------------------------------------
(18) 19th April 1995; bug reported by Kim Horn
For very small values of CF less than 0.1%, confidence levels
are computed erratically. Modified stats.c.
--------------------------------------------------------------------------------
(19) June 1995: modifications to siftrules.c (again!)
Scheme described above in 14(a) amended in line with my ML'95
paper, available by anonymous ftp from ftp.cs.su.oz.au, directory
pub/ml, file q.ml95.ps.Z.
--------------------------------------------------------------------------------
*********************** Release 7 June 1995 *********************************
--------------------------------------------------------------------------------
(20) 6th July 1995; bug reported by Andrew Taylor
Tree printing can have problems when attribute names are very long.
Modified trees.c.
--------------------------------------------------------------------------------
(21) 18th October 1995: modifications to contin.c
Altered the calculation of gain for continuous attributes (described
in "Improved Use of Continuous Attributes in C4.5").
--------------------------------------------------------------------------------
*********************** Release 8 October 1995 ******************************
--------------------------------------------------------------------------------
(22) 26th Feb 1996; minor glitches reported by Ron Kohavi of SGI
Fn declared extern in rules.c
-lm removed from consult, consultr, xval-prep
--------------------------------------------------------------------------------
