ripperaux.1

Ripper 分类算法

.EN
.TH RIPPER AUXILIARY PROGRAMS 1
.SH SYNOPSIS
.PP
data2text \- convert feature-vector data to a single text-valued
field.
.PP
eliminate \- eliminate obviously redundant conditions from a ruleset
.PP
filter-text \- perform feature selection on text, optionally converting
it to an attribute-value format.
.PP
pair \- do a paired test
.PP 
partition \- randomly subsample a dataset
.PP
pprint-rules \- print a ruleset
.PP
predict \- makes predictions using a ripper-generated ruleset 
.PP
rocchio \- strawman text categorizer
.PP
select-class \- remove all but two classes from a dataset
.PP
summarize \- summarize output of predict.
.PP
test-rules \- test individual ripper rules on data
.PP
verify \- check a dataset for consistency
.SH DESCRIPTION
Miscellaneous programs to assist in evaluating the rulesets
produced by RIPPER.  All of these programs will summarize
their legal options if invoked with a -h option. 
Most of them also will read from standard input if invoked
with a -s option. 
.SH PREDICT
.B predict [options] filestem
.PP
Uses the ruleset filestem.hyp to predict classes of the examples
in the file filestem.test.  (If the -s option is used then
examples are taken from standard input instead of filestem.test).
Output is a series of lines of the form
.PP
predicted-class p n actual-class 
.PP
where p and n are the numbers of positive and negative examples
classified by the rule used to pick the predicted-class.
.SH SUMMARIZE
.B summarize [options] filestem
.PP
Summarize the predictions appearing on the standard input.
Reports the error rate, recall, precision, and a confusion matrix.
.SH VERIFY
.B verify [options] filestem
.PP 
Do a consistency check on the dataset filestem.data
.SH ELIMINATE
.B eliminate [options] filestem
.PP 
Eliminate redundant conditions from the ruleset filestem.hyp
.SH TEST-RULES
.B test-rules [options] filestem
.PP 
Gives a rule-by-rule breakdown of error rates 
of the rules in filestem.hyp on the examples in filestem.test
.SH PAIR
.B pair predfile-1 predfile-2
.PP 
Does a paired test of the classifiers used to generate the
predictions in the files predfile-1 predfile-2.  Predfile-1
and predfile-2 should be the outputs of predict.  Outputs
the following information: number of "wins" (times the classifier
used for predfile-1 is right and the classifier of predfile-2
is wrong); number of "losses"; number of "ties" (times both
classifiers agree); estimate of the probability of a win
given a disagreement, and the standard error of the estimate;
and how far the probability of winning is from 0.5 
in standard units (standard errors).  
.PP
Standard errors are computed using a normal approximation
to the binomial distribution.
Generally, much more than 2 standard
units means that the difference is statistically significant. 
.SH PARTITION
.B partition [-p q | -m N] [stem]
.PP 
Chooses a random subset of stem.data.  Size of subset
is specified by either a percentage to retain (-p) or by 
a subset size (-m).
.SH SELECT-CLASS
.B select-class -c C [stem]
.PP 
For examples in stem.data, changes class of examples of 
anything other than class C to 'default_class', and writes 
the modified dataset to the standard output. 
.SH PPRINT-RULES
.B pprint-rules [options] stem 
.PP 
Print a RIPPER ruleset to standard output in a human-readable
format.
.SH EXAMPLES
.PP 
Here are some common ways of evaluating RIPPER. 
.PP 
#generate a ruleset 
.PP 
ripper foo 
.PP 
#eliminate redundant conditions
.PP 
eliminate foo 
.PP 
#test ruleset on foo.test and summarize
.PP 
predict foo | summarize foo
.PP
#save predictions in a file
.PP 
predict foo > foo-default.pred
.PP
#re-run experiment with different RIPPER options 
.PP 
ripper -L2 foo
.PP 
predict foo > foo-L2.pred
.PP 
#do a paired test
.PP 
pair foo-default.pred foo-L2.pred
.SH BUGS
Probably.