
A Support Vector Approach to Censored targets:

Authors:  Pannagadatta Shivaswamy
          Wei Chu
          Martin Jansche

Disclaimer:


This software is written to do model selection by cross 
validation and to predict on the test examples.

Instructions for compiling:
--------------------------
There are two ways in which the code can be compiled.

If cross-validation needs to be done on AAE, then use
"make aae". The resulting executable is  str_aae. 

If cross-validation needs to be done on the rankscore, 
use "make rank".  The resulting executable is str_sp.

Instruction for running:
-----------------------
Typing ./str_aae (or str_sp) lists the different options available 
for running the code. 

Training File:
---------------------------
Name: <prefix>_train

Each line of the file represents one example; Each line looks
as below:
<lower bound> <upper bound> <feature_number>:<feature_value> <feature_number:feature_value> <feature_number> <feature_value>

<lower bound> and <upper bound> are the two target values. When they
are equal, the example has a single real value targets; when they are
different, the target becomes an interval. Censored targets can be
represented by either a high value for the upper bound or a very small
(negative) value for the lower bound.

feature_number starts from 1, for nth feature number the corresponding 
feature value represents nth co-variate for that example. Only non zero
features need to be listed; others are assumed to be zero.

Test File:
-----------------------
Name:  <same prefix as train file>_test

Same as the training file. If the targets are unknown, just fill zeros
for the upper and lower bounds of each example.

Output:
-----
Once the program is run, the code does model selection using
cross validation and trains on the entire training set using
the best model. The predictions on the test set are obtained
in <same prefix as train file>_test.guess.

Example train and test files can be found with the package:
toy_train, toy_test.










