Databases Reference
In-Depth Information
Currently in VisMiner, the SVM modeler is a black box. It does not support
analyst tweaking of the parameters. By default, VisMiner uses the radial basis
function. During model construction, with a small subset of the data, it builds
multiple models using varying values of the cost parameter, choosing the best,
then applying to the full dataset. The analyst only sees this final model.
In choosing to build an SVM classifier, another consideration is that of
processing time. SVM algorithms are quite CPU intensive. Typically they will
take much longer to complete than decision tree or ANN processors. You may
want to build using smaller datasets or fewer predictor variables.
Summary
VisMiner implements algorithms for both classification and regression. These
modelers can be deployed by the analyst without having knowledge of the
algorithm internals. However, having that model understanding can result in
better model application decisions and less ambiguous interpretation of the
results.
A problem common to all prediction modelers is the overfitting of the model
to the training data. Each modeler has guidelines and processes that can be
followed to avoid this problem. A knowledge of the algorithm internals helps in
their application.
A summary of the advantages and limitation of each algorithm is found in
Appendix A, Table A.1.
