Precision and recall are a great way of looking at overall accuracy, but ROC curves and ROC AUC are highly recommended for model selection criteria on imbalnced datasets. For SVMs you can do a two dimensional grid search for C and Gamma looking for a maximum ROC.
I also use per class weights - by default SVM(libSVM, libLinear) cost parameter is the same for both classes. Penalize the classifier for false negatives more then for false positives(order of magnitude more).
I also use per class weights - by default SVM(libSVM, libLinear) cost parameter is the same for both classes. Penalize the classifier for false negatives more then for false positives(order of magnitude more).