We did an initial wide grid search over cost and gamma, then focused on a gamma from 0 to 0.05 and a cost of 0 to 40. Cost penalizes misclassifying training examples. Gamma is a parameter of the kernel which allows less computationally expensive operation in high dimensional space [keep this explanation?] Gamma is recommended to have a value of 1 / Amount of features = 0.0002757 after SMOTE preprocessing. However, we get better results with higher values. Cost is recommended to be modified by factors of ten, but we focused on the range in the graph above as it seemed sufficiently promising. It is possible that we have found a local maximum or even that the model now overfits the data. The highest kappa value of 0.82 in this search was found at a Cost of 10 and a Gamma of 0.01. This highest point seems to be in a stable region.

Because of the good result and high speed of the LibSVM method, we recommend this method.