GBT-SVM

In this repository, a MATLAB toolbox is provided to train and test the support vector machine (SVM) based models for imbalanced and multi-type classification problem. Two classification tricks namely granularization and binary-tree have been appended to form the GBT-SVM model. Please refer to my paper for details of the model.

Construction of the toolbox

Codes and scripts in the code folder can be used to build your granular SVM (GSVM) models, which are able to form the tree-structure classifiers. I list the name and the corresponding comment here.

Method	Comment
getGranule	Generate balanced datasets by spitting the major class into subsets or granules.
myCrossSVM	Train SVM models with cross-validation and gridding, kernels are optional.
myGSVMpredict	Classify on new samples with the trained model.

Get granules

To get balanced dataset, you may use getGranule as

>>> [DataGranules] = getGranule(data,label)

The output DataGranules is a structure, which is formed as,

DataGranular

• MajIdx: label of the major class
• MinIdx: label of the minor class
• GraNum: number of granules
• MinData: data and labels of the minor-class samples
• MajData: data and labels of the major-class samples
• GraIdx: indices of the major samples that are split into the granules

Train a granule with SVM

To train a SVM model, please use the myCrossSVM method,

>>> [model,normMat] = myCrossSVM(data,label,'RBF');

where 'RBF' is the radical basis function kernel, other kernel like linear can be selected.The output normMat keeps the column-wise minimum and maximum values to normalize the new samples as the same as the samples for training. The normalization is a very important issue in machine-learning tasks.

Example

An example is demonstrated to use the GBT-SVM toolbox, please refer to GSVM_train.m for details. The sample set and trained models are also included in this repo for a quick start. To evaluate the models and classify a new sample by the GSVM under a voting strategy, please refer to GSVM_test.m.

If you want to train the GSVM models with multi-threads, please use the codes as follows,

parpool(2) % Two threads
parfor i = 1 : GraNum
    SubSetMaj = DataGranules.MajData.SampleSet(DataGranules.GraIdx{i},:);
    SubLabelMaj = DataGranules.MajIdx * ones(length(DataGranules.GraIdx{i}),1);
    SubTrainSet = [SubSetMaj;SubSetMin];
    SubTrainLabel = [SubLabelMaj;SubLabelMin];
    [ModelGSVM{i}.model,ModelGSVM{i}.PS] = myCrossSVM(SubTrainSet,SubTrainLabel,'RBF');
end

Requirements

To use the toolbox, the libsvm is required. Before using the methods in the libsvm, you may reconfigure the library with system based compiler, e.g., the VC++ for windows, and the g++ for linux.

Please follow the steps to install the libsvm (my libsvm version is 3.18)

1. Unzip and put 'libsvm-3.18' into the path xxx/matlab/toolbox
2. Add the path of xxx/matlab/toolbox/libsvm-3.18/matlab/ to the matlab environment, you would like to use the Set Path button in the Home panel. Remember to save the change.
3. Reconfigure the library with matlab a. In the Command Window, input mex -setup to choose a compiler; b. Check the compiler, and configure. Input make to begin the compilation.

Citation

This work has been published in ICSP 2016, which can be cited as following bibtex,

@InProceedings{Ma2016,
    author = {Ma, Z. and Wang, L. and Li, W. and Xu, H. and Zhu, J.},
    title  = {X-ray astronomical point sources recognition using granular binary-tree svm},
    booktitle = {Signal Processing (ICSP), 2016 13th IEEE International Conference on},
    year = {2016},
    pages = {1021--1026},
    address = {Chengdu, China},
    month   = nov,
    organization = {IEEE},
    doi = {10.1109/ICSP.2016.7877984},
    url = {http://ieeexplore.ieee.org/document/7877984/}
    }