It is written in C/C++ but has also Java source code. The learning tasks are 1) support vector classification (binary and multi-class), 2) support vector regression, and 3) distribution estimation. Supported problem formulation are: C-Support Vector Classification, ν-Support Vector Classification, distribution estimation (one-class SVM), ɛ-Support Vector Regression, and νSupport Vector Regression. All of the formulations are quadratic minimization problems and are solved by sequential minimal optimization algorithm. The running time of minimizing SVM quadratic problems is reduced by shrinking and caching. LibSVM provides some special setting for unbalanced data by using different penalty parameters in the SVM problem formulation. It was successfully used in computer vision, NLP, neuro-imaging, and bioinformatics (since 2000 to 2010 with 250 000 downloads). It is also included in some DM environments: RapidMiner, PCP, and LIONsolver. The SVM learning code from the library is often reused in other open source ML toolkits, including GATE [Gate], KNIME [Knime], Orange [Orange] and scikit-learn. The library is very popular at open source ML community (released under the 3-clause BSD license). LibSVM version 3.22 released on December, 2016.

Strong points

• The LibSVM data format is a specific data format for the data analysis tool LibSVM, which is well-accepted in other frameworks and libraries. The format is dense and suitable to describe and process Big Data especially because it allows for a sparse representation.
• Open source, well-maintained and specialised tool with high popularity in open source ML community.

Weak points

• LibSVM training algorithm does not scale up well for very large datasets in comparison with LibLinear or Vowpal Wabbit [Zygmunt 2014]. It takes O(n³) time in the worst case and around O(n²) on typical cases.
• Limited to problems, with which SVM deals well.

Return to Contemt