BudgetedSVM - Version 1.1

The toolbox is distributed as an open-source software. However, before downloading, please make yourself familiar with the Modified BSD license.
To download BudgetedSVM please click here. Note that you will be redirected to SourceForge.net website, and the download will start after approximately 5 seconds.

If you would like to contribute to the project by helping us make BudgetedSVM better, faster, and more scalable piece of software, please check our BudgetedSVM project page at SourceForge.net. Any contributions on your side, such as comments, bug reports, praises, etc., are more than welcome. For a more direct communication, you are also welcome to send an e-mail to nemanja(at)temple(dot)edu.

Lastly, if you find the toolbox useful, please cite our JMLR paper.

Earlier versions of the software can be found here:

Data sets and LibSVM format

BudgetedSVM toolbox works with data sets stored in .txt files in LibSVM sparse-data format. The format stores each data point in one line, starting with the label at the start of the line, and followed by feature_index:feature_value pairs separated by a space, where indexing starts from 1. If a certain value of feature_index does not appear in the line, this indicates that the feature with that index is equal to 0. For example, the first three lines of a9a_test.txt file, storing the testing examples of adult9a data set, are given here

-1 1:1 6:1 17:1 21:1 35:1 42:1 54:1 62:1 71:1 73:1 74:1 76:1 80:1 83:1
-1 3:1 6:1 14:1 22:1 36:1 40:1 56:1 63:1 67:1 73:1 74:1 76:1 82:1 83:1
+1 2:1 10:1 18:1 24:1 38:1 40:1 59:1 63:1 67:1 73:1 74:1 76:1 80:1 83:1

We can see that the first example is of the negative class, with features indexed by 1, 6, 17, 21, etc., equal to 1, while the remaining features are equal to 0. Similarly, the third example is of the positive class, with features indexed by 2, 10, 18, 24, etc., equal to 1, while the remaining features are equal to 0. For more details, as well as a large collection of standard and less-standard benchmark data sets in this format, please check LibSVM website.

Known issues

If you experience problems with BudgetedSVM, please let us know so we can address a potential issue as soon as possible. The following is the list of known issues with the package:

  • We experienced problems on Mac machines, which resulted in a trained model not being loaded properly from a model file. Although it is possible that the problem can be replicated on different Macs, we noticed it on OS X, version 10.9.1, with gcc version 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.11.00). If you are using Mac, please download the following zip file, and overwrite the files in "src" directory with the included .cpp source files.


The toolbox was written by Nemanja Djuric, Liang Lan, and Slobodan Vucetic from the Department of Computer and Information Sciences, Temple University, together with Zhuang Wang from IBM, Global Business Services - Business Analytics and Optimization.

We acknowledge that this work was supported in part by the National Science Foundation via the grant NSF-IIS-0546155.