sba1 / clperf Goto Github PK

clperf - evaluating classifier performance
==========================================

The program clperf is a simple command line tool that
can be used to evaluate the performance of classification
procedures. As input it accepts an ASCII file that contains
tab-separated values (TSV) in which one column identifes the
true label, and another column stores the actual prediction
value produced by the classifer. The purpose of clperf
is then to calculate measures such as ROC or Precision/Recall
on the input.

While there are many extensions for statistcal software
available that follow a similar goal, clpref has some
interesting features such as:

 - employs external memory algorithms and thus allows to
   handle large input files (beyond capablities of the
   computers physical memory) can be run
 - it is a command line tool and thus easily scriptable
 - samples the resulting graph in an equi-distance manner
 - the result can be written in various formats for further
   processing
 - code is written in portable ISO C99

Building
========

In order to build clperf, simply type
 make
after the command prompt. The executable is generated
in the same directory.


Usage
=====

 clperf [OPTION] INPUT LABELCOL PREDCOL

INPUT is the file name of the TSV
LABELCOL is the column number (0-based) of the label
     The label is either a 0 (negative class) or 1 (positive
     class)
PREDCOL is the column number (0-based) of the prediction
     a negative column number indicates that the order is
     reversed

Currently, clperf writes an R script to the stdout that, when
invoked within R, draws a ROC and Precision/Recall plot. Note
that this may change in the future. Refer to the built-in
help that can be seen via the --help option to learn more
about possible options and their effect.


Contact
=======

You can reach the author by writting an email to
 [email protected]


License
=======

The program clperf is written by Sebastian Bauer and released
under the conditions of the AGPL. See the License file for
more details.