cssag
=====

An implementaion of a Multi-Label Classification algorithm on Tree- and DAG-Structured Label Hierarchies

See: http://machinelearning.wustl.edu/mlpapers/paper_files/ICML2011Bi_10.pdf for detailed description of the algo.


Usage: 
======
hierarchicalMultiLabelLearning.py [options]

Options:
  -h, --help            show this help message and exit
  -i HIERARCHY, --hierarchy=HIERARCHY
                        the file with the label hierarchy information: Each
                        line in the file is <nodeID> <childNodeID1>
                        <childNodeID2>... <childNodeIDn>
  -e TESTFILE, --test=TESTFILE
                        the file for testing in svmlight format
  -a TRAINFILE, --train=TRAINFILE
                        the file for training in svmlight format
  -x TESTLABELS, --test-labels=TESTLABELS
                        the test labels file: comma separated labels per line
  -y TRAINLABELS, --train-labels=TRAINLABELS
                        the train labels file: comma separated labels per line
  -d DIM, --dim=DIM     number of dimensions in the PCA
  -t TREE, --tree=TREE  selest AND or OR tree: takes input "and" or "or"
  -r REGRESSOR, --regressor=REGRESSOR
                        type of regressor to use: "ridge" or "svr"
  -p PCA, --pca=PCA     type of pca: "pca"/"kpca"(kernel PCA)/"nopca"(a
                        regressor is trained on each label)

For example, using the train, test and hierarchy files in this folder:

    hierarchicalMultiLabelLearning.py -i hierarchy.dat -e test.svmlight.mini -a train.svmlight.mini -x test.labels.mini -y train.labels.mini 

Notice that these options are the compulsory options, while dimension, pca, tree type and regressor type are optional and if not provided, default values will be used. An example of a complete command:

    hierarchicalMultiLabelLearning.py -i hierarchy.dat -e test.svmlight.mini -a train.svmlight.mini -x test.labels.mini -y train.labels.mini -d 100 -t or -r svr -p pca


Input file format:
====================
The train and test files are in svmlight format. Since it's a multilabel classification problem, the first number in each line(which usually denotes the class the sample belongs to) is meaningless in the svmlight file: so you are needed to give the label information per sample in a corresponding label file, the line n of which represents labels that describe the sample n (or line n) in the corresponding svmlight file. The labels are comma separated.

The hierachy file defines the label hierarchy. Each line has information on a node and its children, which are space separated: <nodeID> <childID1> ... <childIDn> . A line corresponding to a node with no children is: <nodeID> . 

See included example files.

Other options:
==============
Please see the original paper for these variables:

DIM defines the dimension of the PCA that maps the trainspace to a lower dimensional space.

The TREE option defines whether the inherent hierarchy conforms to the AND-G property or the OR-G property defined in the paper.

Regressor to train the DIM number of reduced dimensions: can be either ridge regression or support-vector regresson

PCA: If you want to follow the paper religiously, then 'kpca' option is what you are looking for. To compare the results to other possible modifications, I have also included 'pca' and 'nopca' options. The 'pca' option follows Tai & Lin(2010)(as mentioned in this paper) and performs PCA directly on label vectors where are the 'kpca' uses kernel PCA on a transformed feature vector set. Using 'nopca' would mean no transformation to lower dimension is implemented and a regressor is trained for each label in the hierarchy.