The ambition of this master thesis is to generate an R package of the existing program called potential support vector machine (PSVM). Currently only a version under C++ and MATLAB exists. The advantage of the PSVM compared to standard SVMs is that the PSVM is invariant under linear transformations. These would be for example the scaling of the data. Also dyadic data representation is included in the PSVM. Another important point is, that the program can also handle non positive kernel functions. The present program includes methods for training a model and predicting values for new data. Additionally the functionality should be extended. In the future the usage of multi-class data will be available. In the correlation with this extension, the functionality of probability estimates for the particular classes will be included. This based on the algorithm of the so called Platt scaling.
Forschungs-